WO2022255332A1 - Method for screening for retinopathy of prematurity, apparatus for screening for retinopathy of prematurity, and learned model - Google Patents

Method for screening for retinopathy of prematurity, apparatus for screening for retinopathy of prematurity, and learned model Download PDF

Info

Publication number
WO2022255332A1
WO2022255332A1 PCT/JP2022/022042 JP2022022042W WO2022255332A1 WO 2022255332 A1 WO2022255332 A1 WO 2022255332A1 JP 2022022042 W JP2022022042 W JP 2022022042W WO 2022255332 A1 WO2022255332 A1 WO 2022255332A1
Authority
WO
WIPO (PCT)
Prior art keywords
treatment
retinopathy
prematurity
progression
screening
Prior art date
Application number
PCT/JP2022/022042
Other languages
French (fr)
Japanese (ja)
Inventor
福嶋葉子
西田幸二
小島諒介
奥野恭史
Original Assignee
国立大学法人大阪大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 国立大学法人大阪大学 filed Critical 国立大学法人大阪大学
Priority to US18/566,168 priority Critical patent/US20240274295A1/en
Priority to JP2023525837A priority patent/JPWO2022255332A1/ja
Publication of WO2022255332A1 publication Critical patent/WO2022255332A1/en

Links

Images

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/30—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B10/00—Instruments for taking body samples for diagnostic purposes; Other methods or instruments for diagnosis, e.g. for vaccination diagnosis, sex determination or ovulation-period determination; Throat striking implements
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H20/00—ICT specially adapted for therapies or health-improving plans, e.g. for handling prescriptions, for steering therapy or for monitoring patient compliance
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems

Definitions

  • the present disclosure relates to a screening method for retinopathy of prematurity, a screening device, and a trained model for predicting the progression of retinopathy of prematurity, which are used to assist doctors in diagnosing retinopathy of prematurity.
  • Retinopathy of prematurity is a major cause of blindness in childhood, and it is estimated that 50,000 people worldwide become blind each year, and the number is expected to increase further in the future. In many cases, spontaneous remission can be expected after the onset of the disease, but in immature, extremely premature infants, intraocular hemorrhage and retinal detachment may lead to blindness. Retinal photocoagulation is the standard treatment for such severe cases. Although retinal photocoagulation has been shown to be effective in suppressing the progression of retinopathy of prematurity, it avoids blindness at the expense of tissue destruction, and is not a preventative treatment.
  • a method for screening for retinopathy of prematurity includes, for example, the technology described in Patent Document 1.
  • tryptase which can be released by degranulation of mast cells, is detected as a marker substance from the blood derived from the subject to determine whether treatment for retinopathy of prematurity is necessary.
  • WINROP a model developed in Sweden
  • CHOP-ROP model a model reported from the United States
  • WINROP targets gestational age of 23 weeks or more and less than 32 weeks, and if the gestational age, birth weight, and postnatal weight are entered every other week, an alarm will be set for cases where there is a possibility of deterioration. Is displayed.
  • the CHOP-ROP model like WINROP, can be evaluated by weekly weight gain and can reduce the number of consultations in the low-risk group.
  • Patent Document 1 requires an invasive means of blood sampling, which is not realistic.
  • the treatment of retinopathy of prematurity is necessary by the method described in Patent Document 1, it may be cured spontaneously, and even if it is determined that the treatment of retinopathy of prematurity is unnecessary, it takes several days. Later, retinopathy of prematurity may develop and become severe. For this reason, frequent fundus examinations are important for predicting the progression of retinopathy of prematurity.
  • One aspect of the present disclosure is a screening method for retinopathy of prematurity for predicting the progression of retinopathy of prematurity, comprising a postnatal time series of weight, height, and vital signs of a premature infant whose gestational age is less than a predetermined week.
  • the method includes a treatment determination step for determining whether or not treatment for retinopathy of prematurity is indicated after a predetermined number of days after birth, based on premature infant information including data.
  • a screening device that predicts the progression of retinopathy of prematurity, and includes postnatal chronological data on the weight, height, and vital signs of a premature infant whose gestational age is less than a predetermined week.
  • the present invention is characterized by including a treatment determination unit that determines whether or not treatment for retinopathy of prematurity is indicated after a predetermined number of days after birth, based on premature infant information.
  • the progression of retinopathy of prematurity is predicted by determining whether treatment for retinopathy of prematurity is indicated using postnatal time-series data on weight, height, and vital signs. be able to. Many factors are intricately related to the progression of retinopathy of prematurity from onset to treatment indication, and these factors change over time.
  • This configuration provides a method or apparatus for predicting the progression of retinopathy of prematurity with high accuracy using a plurality of time-series data.
  • This method or this device uses information including changes in immaturity and general condition over time. For example, the gestational age at birth is used as an index for immaturity, and the general condition after birth is used as an index for time-series data including vital signs such as heart rate, respiration and blood oxygen concentration, as well as body weight and height.
  • it is determined whether or not treatment for retinopathy of prematurity is indicated after a predetermined number of days after birth.
  • a program is installed in equipment for chronological monitoring of vital signs of premature infants in the neonatal intensive care unit, and in the near future, retinopathy of prematurity will occur. These include diagnostic aids that provide warning signs for treatment-indicated cases.
  • the progression of retinopathy of prematurity can be predicted with higher accuracy than existing models (WINROP and CHOP-ROP models).
  • Timely and appropriate treatment of retinopathy of prematurity reduces the risk of blindness.
  • the shortage of ophthalmologists skilled in accurate staging and treatment has become a serious problem worldwide.
  • the stage of retinopathy of prematurity is determined from images obtained by ophthalmologists using an indirect ocular fundus examination or a contact-type fundus camera.
  • the method or device determines suitability for treatment based on preterm infant information, including postnatal chronological data on weight, height, and vital signs of preterm infants, without the need for fundus examination or imaging. This also helps reduce unnecessary medical examinations. In other words, it not only contributes to medical care, but also contributes to society and the economy by significantly reducing medical costs due to visual impairment, lower productivity, and costs related to social care.
  • Another embodiment of this method further comprises a risk determination step of determining the risk of progression of retinopathy of prematurity based on the Type 1 ROP or APROP score calculated at the predetermined number of days after birth, wherein the The point is that the treatment determination step is performed only for the premature infant determined to have a progression risk.
  • another aspect of the present device further comprises a risk determination unit that determines the risk of progression of retinopathy of prematurity based on the Type 1 ROP or APROP score calculated at the predetermined number of days after birth, and the treatment determination unit The point is to determine whether or not the treatment is indicated for only the premature infant determined by the risk determination unit to have the progression risk.
  • the latent degree of progression of retinopathy of prematurity can be estimated in order to determine the risk of progression of retinopathy of prematurity.
  • the two-step determination process provides a highly accurate screening method or screening device for retinopathy of prematurity.
  • the vital sign is at least one of heart rate, respiratory rate and arterial blood oxygen saturation of the premature infant.
  • the preterm infant information includes at least one of the gestational age and Apgar score of the preterm infant.
  • the progression of retinopathy of prematurity can be predicted more accurately by using the gestational age and Apgar score of premature infants in addition to the postnatal general condition as premature infant information.
  • One aspect of the present disclosure is a trained model that functions by a computer, is composed of a decision tree consisting of a plurality of branch points arranged in a tree structure, and has a predetermined number of weeks of gestation when treatment for retinopathy of prematurity is performed.
  • a feature value calculated based on preterm infant information including postnatal time-series data on the weight, height and vital signs of a preterm infant that is less than is input, and by summing the evaluation values at each of the branch points, The point is to output a score indicating the necessity of treatment for retinopathy of prematurity.
  • One aspect of the present disclosure is a computer-functioning trained model, generated by deep learning including a convolutional neural network, for premature infants treated for retinopathy of prematurity with a gestational age of less than a predetermined number of weeks.
  • Premature infant information including postnatal chronological data on weight, height and vital signs is input, and a score for the need for treatment of retinopathy of prematurity is output.
  • a trained model that has undergone deep learning, including a convolutional neural network, as in this configuration, is highly versatile and can accurately predict the progression of retinopathy of prematurity at an appropriate timing.
  • the premature infant information is information obtained from the premature infant whose total gestational age and number of weeks at the time of treatment is 40 weeks or less.
  • the premature infant information excludes information obtained from the premature infants who were treated early based on the doctor's judgment.
  • the vital sign is at least one of heart rate, respiratory rate and arterial blood oxygen saturation of the premature infant.
  • Such vital signs are acquired by existing monitoring equipment installed in the neonatal intensive care unit, so it is possible to secure a large amount of input data for building a trained model.
  • the preterm infant information includes at least one of the gestational age and Apgar score of the preterm infant.
  • FIG. 1 is an overall view of a system for realizing a screening method according to this embodiment.
  • FIG. 1 is a block diagram of a screening device according to this embodiment;
  • FIG. 1] is a flowchart for realizing a screening method according to the present embodiment.
  • FIG. 1] is an explanatory diagram of a screening method according to the present embodiment.
  • FIG. 10 is a diagram showing the relationship between the number of weeks at the time of treatment or the gestational age and the total number of weeks at the time of treatment and the treatment results. is a diagram showing an example of a risk determination process.
  • FIG. 4 is a diagram showing the degree of contribution of each feature amount in machine learning
  • FIG. 10 is an example ROC curve diagram in which a treatment decision process is performed using a machine-learned trained model
  • FIG. 4 is an AUC diagram of an example of performing a treatment determination process using a trained model that has undergone deep learning
  • One or more monitoring devices 1 for obtaining preterm infant information are connected to the Internet line 2.
  • the Internet line 2 is connected with a trained model generation device 3, a screening device 4, and an AI 9 (artificial intelligence).
  • the AI 9 may be provided on the Internet line 2 or may be provided in the trained model generation device 3 .
  • the trained model generation device 3 and the screening device 4 may be the same device, or the screening device 4 may be incorporated in the monitoring device 1, and their functions may be used singly or in combination. can.
  • the screening device 4 may be a monitor device installed in an incubator or a dedicated device in a neonatal intensive care unit, and can be used as various diagnostic auxiliary devices for predicting the progression of retinopathy of prematurity.
  • the monitoring device 1 is a device that monitors the vital signs of preterm infants over time and a device that periodically measures the weight and height of preterm infants in the neonatal intensive care unit.
  • the preterm infant information of this embodiment includes the weight, height, and vital signs of a preterm infant whose gestational age is less than a predetermined number of weeks (for example, 28 weeks).
  • This predetermined week is 36 weeks or less (so-called premature infants), preferably 32 weeks or less, more preferably 28 weeks or less (so-called very premature infants) (hereinafter the same).
  • birth weight may also be used as an index.
  • This birth weight is less than 2500 g (so-called low birth weight infants), preferably less than 1500 g (so-called very low birth weight infants), more preferably less than 1000 g (so-called very low birth weight infants).
  • Vital signs are at least one of heart rate, respiratory rate and arterial oxygen saturation of premature infants and are obtained at least every minute. The vital signs may include blood pressure and the like as long as they are vital information of the premature infant.
  • Premature infant information includes weight of preterm infants obtained three times a week, height of preterm infants obtained once a week, gestational age of preterm infants, and 1 and 5 minutes after birth. It preferably contains at least one of the later Apgar scores.
  • the acquisition frequency of premature infant information is not particularly limited, and may be set every second or every 10 minutes for vital signs, or every day for weight and height.
  • the Apgar score evaluates the condition of a newborn immediately after birth, and evaluates the five items of skin color, heart rate, reaction, muscle tone, and respiration on a total of 10 points.
  • the trained model generation device 3 includes a first communication unit 31, a model generation unit 32, a learning feature value calculation unit 33, and a first storage unit .
  • the first communication unit 31 is an interface that transmits and receives data to and from the monitoring device 1, the screening device 4, the AI 9, etc. via the Internet line 2.
  • the first communication unit 31 may receive data directly from the monitoring device 1, or accumulate data acquired by the monitoring device 1 in a server (not shown) and receive the accumulated data from this server. Also good.
  • the first storage unit 34 is composed of a non-temporary storage medium such as an HDD or SSD or a temporary storage medium such as a RAM, and stores programs and applications to be executed by the processor.
  • the first storage unit 34 stores learning premature infant information 34 a of the monitoring device 1 acquired via the first communication unit 31 .
  • the learning preterm infant information 34a includes postnatal chronological data on the weight, height, and vital signs of preterm infants whose gestational age is less than a predetermined week (for example, 28 weeks).
  • the premature infant information for learning 34a includes the gestational age and Apgar score of the premature infant.
  • This premature infant information for learning 34a is preferably information obtained from a premature infant whose total gestational age and number of weeks at the time of treatment is 40 weeks or less.
  • the premature infant information for learning 34a excludes information obtained from premature infants who were treated early by doctor's judgment.
  • the first storage unit 34 stores treatment information 34c associated with learning premature infant information 34a.
  • This treatment information 34c is classified into no treatment, Type 1 ROP with treatment, and APROP (Aggressive Posterior Retinopathy Of Prematurity) with treatment.
  • the International Classification of Classic ROP (typical ROP) for retinopathy of prematurity includes Type 1 ROP and Type 2 ROP, with Type 1 being indicated for treatment and Type 2 being less than indicated for treatment (no treatment).
  • APROP a type that rapidly worsens apart from Classic ROP is called APROP, which is also indicated for treatment.
  • Type 1 ROP international classification
  • APROP international classification
  • treatment indication means to implement treatment within 72 hours after being diagnosed with Type 1 ROP, or to implement treatment promptly if there is an early sign of APROP (hereinafter the same).
  • Type 1 ROP with indications for treatment is either zone1 ROP with plus disease, zone1 stage3 ROP without plus disease, or zone2 stage3 ROP with plus disease, or APROP.
  • zone 1 is the area within a circle centered on the optic nerve head with a radius twice the distance between the head and the macula
  • zone 2 is the area within a circle with a radius from the papilla to the nasal serrated margin
  • stage 3 is extraretinal fibrovascular proliferation.
  • the first storage unit 34 stores the trained model 10.
  • the trained model 10 is a model that functions by a computer and is obtained by machine learning or deep learning with supervised data. Further, the first storage unit 34 stores a learning feature quantity 34b calculated by the learning feature quantity calculation unit 33 for machine learning.
  • Machine learning consists of a decision tree consisting of multiple branch points arranged in a tree structure. This machine learning has a structure in which a feature amount is evaluated at each branch point of a decision tree, and an evaluation value corresponding to the evaluation result is given to each branch point. Then, the evaluation values are summed up along the branches of the decision tree to obtain progression prediction information for retinopathy of prematurity.
  • Machine learning may be performed using an ensemble model such as XGBoost, Random Forest, LightGBM, CatBoost, AdaBoost, etc., in which a plurality of decision trees are associated.
  • Deep learning is performed by AI9, including well-known convolutional neural networks (CNN, DCGAN, etc.).
  • CNN convolutional neural networks
  • DCGAN digital signal processor
  • a convolutional neural network constructs a deep-hierarchical model imitating a human neural circuit and infers the progression prediction of retinopathy of prematurity.
  • This deep learning consists of known applications provided via the Internet line 2 .
  • the model generation unit 32 includes a processor and generates the trained model 10.
  • the processor includes ASIC, FGPA, CPU, or other hardware for executing applications or the like stored in the first storage unit 34 (the same applies hereinafter).
  • the model generation unit 32 performs reinforcement learning with the input data as the learning feature amount 34b and the teacher data as the treatment information 34c (no treatment, Type 1 ROP with treatment, APROP with treatment). to generate a trained model 10.
  • the feature value for learning 34b is calculated by processing the premature infant information for learning 34a (time-series data after birth regarding the weight, height and vital signs of the premature infant, the gestational age and Apgar score of the premature infant, etc.). Details will be described later.
  • the model generating unit 32 uses the input data as premature infant information 34a for learning (time-series data after birth regarding the weight, height and vital signs of the premature infant, etc.), Reinforcement learning is performed with teacher data as treatment information 34c (no treatment, Type 1 ROP with treatment, APROP with treatment), and a learned model 10 is generated.
  • the learning feature quantity calculation unit 33 includes a processor, calculates a plurality of learning feature quantities 34b from the learning premature infant information 34a, and inputs the calculated plurality of learning feature quantities 34b to the trained model 10. Then, the degree of influence of each feature amount on the treatment information 34c is obtained.
  • FIG. 8 illustrates the degree of influence of a plurality of learning feature values 34b on the treatment information 34c.
  • Weight, height, height_SD gestational age, Apgar score 5 minutes after birth (Apgar score; 5min), weight_SD, Apgar score 1 minute after birth (Apgar score 1 min), arterial oxygen saturation (SpO2 %), heart rate (HR bpm), sex (M or F), onset, birth type (single, twin, conception), count, respiratory rate ( RESP/min), respiration rate difference (RESP/min.delta), heart rate difference (HR bpm.delta), weight_SD difference, body weight difference, arterial blood oxygen saturation difference (SpO2 %.delta) , height_SD difference, and height difference.
  • the learning feature quantity calculation unit 33 extracts a plurality of (for example, 10) indices in descending order of influence, and the model generation unit 32 uses the extracted plurality of indices as the learning feature quantity 34b. Also good.
  • Weight is forward-interpolated from the weight of postnatal premature infants obtained three times a week as a daily average (or hourly average).
  • the height is obtained by forward-interpolating the height of the preterm infant after birth obtained once a week as a daily average value (or an hourly average value).
  • _SD is a numerical value called SD (standard deviation) that represents the degree of variation from the average value, that is, the width of the distribution.
  • Arterial oxygen saturation, heart rate, and respiratory rate were interpolated as daily average values (or hourly average values) by removing zero values from the data obtained from preterm infant vital signs after birth every minute. It is.
  • Onset is the finding of retinopathy of prematurity (whether or not retinopathy of prematurity has developed) by a doctor, which is performed a predetermined number of times after birth.
  • a count is obtained by converting the number of days elapsed from the date of birth into a unit time.
  • the difference is a difference feature quantity obtained by calculating the difference for each measurement of each parameter.
  • the weight, height, and vital signs which are the input data in this embodiment, use daily average values, but may be 1-minute average values to 2-day average values, preferably 1-hour average values to 1 day. It is an average value, more preferably an hourly average value or a daily average value (hereinafter the same). If the average value for more than 2 days is used as input data, the prediction accuracy will be poor. .
  • the learning feature quantity calculation unit 33 is built into the trained model 10 .
  • weighting is performed using the time-series feature information inside the trained model 10 (convolution layer), and weighting is performed using the past predicted feature amount inside the trained model 10 (attention mechanism).
  • the time-series feature information is a feature amount obtained by arranging the learning premature infant information 34a in time series
  • the past prediction feature amount is the past time-series feature amount as a result predicted by the trained model 10 itself. It is a weighted feature extracted to use the weighting of feature information for the current prediction.
  • the trained model 10 generated in this way outputs a score indicating the necessity of treatment for retinopathy of prematurity.
  • the need for treatment in this embodiment is classified into no treatment, Type 1 ROP with treatment, and APROP with treatment.
  • Each time series is represented by AUC (Area Under the Curve).
  • This trained model 10 is input with preterm infant information including postnatal chronological data on the weight, height and vital signs of a preterm infant whose gestational age is less than a predetermined week (for example, 28 weeks). It is possible to output scores for each of the postnatal days (for example, 20 days) without treatment, Type 1 ROP with treatment, and APROP with treatment.
  • This predetermined number of days after birth is 1 week or more and 5 weeks or less, preferably 2 weeks or more and 4 weeks or less, and more preferably about 3 weeks (hereinafter the same).
  • Scores are expressed as daily or hourly values (eg, AUC over time). In this embodiment, the score is calculated every day, but the score calculation interval is every minute to every two days, preferably every hour to every day, more preferably every hour or every day ( hereinafter the same). If the score calculation interval is more than two days, the prediction accuracy will be poor, and if it is less than one minute, the amount of data will be large, which may lead to a decrease in calculation speed and noise. This score is calculated for each of no treatment, Type 1 ROP with treatment, and APROP with treatment.
  • the treatment time window means that treatment should be performed within 72 hours in the case of Type 1 ROP with treatment, or that treatment should be promptly performed in the case of APROP with treatment.
  • the treatment is selected from any of retinal photocoagulation, retinal cryocoagulation, intravitreal administration of an anti-VEGF drug, and vitrectomy, preferably retinal photocoagulation or intravitreal administration of an anti-VEGF drug. . Vitrectomy is performed when retinal detachment develops after retinal photocoagulation or anti-VEGF drug therapy is inadequate.
  • the learned model 10 may be configured to output a treatment implementation time or a treatment time range, in addition to expressing as a score whether or not treatment will be performed (treatment indicated) in the future.
  • the screening device 4 includes a second communication unit 41, a prediction feature quantity calculation unit 42, a risk determination unit 43, a treatment determination unit 44, a notification unit 45, and a second storage unit 46.
  • the second communication unit 41 is an interface that transmits and receives data to and from the monitoring device 1, the trained model generation device 3, etc. via the Internet line 2.
  • the second communication unit 41 may receive data directly from the monitoring device 1, or accumulate data acquired by the monitoring device 1 in a server (not shown) and receive the accumulated data from this server. Also good.
  • the second storage unit 46 is composed of a non-temporary storage medium such as an HDD or SSD or a temporary storage medium such as a RAM, and stores programs and applications to be executed by the processor.
  • the second storage unit 46 stores the predicted premature infant information 46 a of the monitoring device 1 acquired via the second communication unit 41 and the learned model 10 generated by the trained model generation device 3 .
  • the predictive preterm infant information 46a includes postnatal chronological data on weight, height and vital signs of preterm infants whose gestational age is less than a predetermined week (eg, 28 weeks).
  • the predictive premature infant information 46a also includes the gestational age and Apgar score of the premature infant.
  • the predictive preterm infant information 46a is preferably information obtained from a preterm infant whose total gestational age and treatment weeks is 40 weeks or less.
  • the second storage unit 46 stores the prediction feature quantity 46b calculated by the prediction feature quantity calculation unit 42 in order to input it to the learned model 10 machine-learned by the learned model generation device 3.
  • the second storage unit 46 stores the determination result 46c output from the trained model 10.
  • This determination result 46c is time-series data divided into no treatment, Type 1 ROP with treatment, and APROP with treatment.
  • the determination result 46c in this embodiment includes ROC (Receiver Operating Characteristic) curves for no treatment, Type 1 ROP with treatment, and APROP with treatment.
  • the determination result 46c also includes a time-series AUC (Area Under the Curve) calculated from this ROC curve.
  • the prediction feature quantity calculation unit 42 includes a processor, and calculates a plurality of prediction feature quantities 46b from the prediction premature infant information 46a.
  • This prediction feature quantity 46b is the same as the learning feature quantity 34b except for onset.
  • the risk determination unit 43 includes a processor, and the trained model 10 to which the predictive premature infant information 46a is input outputs the progression risk of retinopathy of prematurity at a predetermined number of days after birth (for example, 20 days after birth). .
  • Progression risk is expressed as a daily or hourly score. This score is calculated for each of no treatment, Type 1 ROP with treatment, and APROP with treatment. If the value of each score with treatment is higher than a predetermined value at a predetermined number of days after birth, there is a risk of progression. I judge.
  • the risk determination unit 43 extracts the determination index having the highest AUC at a predetermined number of days after birth from the time-series AUC of multiple determination indices composed of no treatment, type 1 ROP with treatment, and APROP with treatment. If the AUC of Type 1 ROP with treatment or APROP with treatment is higher than a predetermined value (for example, 0.3), it is determined that there is a risk of progression.
  • a predetermined value for example, 0.3
  • This predetermined value is set between 0.1 and 0.8, preferably between 0.2 and 0.6, and more preferably between 0.3 and 0.5.
  • the treatment determination unit 44 includes a processor, and predictive preterm infant information 46a including postnatal chronological data on weight, height, and vital signs of preterm infants whose gestational age is less than a predetermined number of weeks (for example, 28 weeks). is input to the trained model 10, it is determined based on the output value of the trained model 10 whether treatment for retinopathy of prematurity is indicated after a predetermined number of days after birth (for example, 20 days). It is preferable that the treatment determination unit 44 determines whether or not only premature infants determined by the risk determination unit 43 to have a risk of progression are indicated for treatment. Whether or not treatment for retinopathy of prematurity is indicated is expressed as a daily or hourly score.
  • This score is calculated for each of no treatment, Type 1 ROP with treatment, and APROP with treatment. It is determined that the treatment time range using is reached (treatment is indicated). As an example, the treatment determination unit 44 extracts the determination index having the highest AUC from time-series AUCs in a plurality of determination indices including no treatment, Type 1 ROP with treatment, and APROP with treatment, and extracts the Type 1 ROP with treatment. Alternatively, if the AUC of APROP with treatment exceeds the AUC without treatment, it is determined that the patient is suitable for treatment.
  • the treatment determination unit 44 determines the AUC of the Type 1 ROP with treatment or the APROP with treatment from the time-series AUC in a plurality of judgment indicators composed of no treatment, Type 1 ROP with treatment, and APROP with treatment. If it exceeds (for example, 0.8), it is determined that the treatment is indicated.
  • This therapeutic threshold is set between 0.5 and 0.9, preferably between 0.6 and 0.9, more preferably between 0.7 and 0.8.
  • the notification unit 45 outputs a warning signal when the treatment determination unit 44 determines that the treatment is indicated.
  • the notification unit 45 may be composed of a warning lamp, a warning sound, or the like mounted on a device for chronologically monitoring the vital signs of premature infants in a neonatal intensive care unit, or a predetermined notification device provided at a nurse station. It may consist of
  • FIGS. 3 to 10 an example of a screening method (program) for retinopathy of prematurity that is executed by a computer to predict the progression of retinopathy of prematurity using the trained model 10 according to the present embodiment. explain.
  • the trained model generation device 3 acquires the learning preterm infant information 34a and the treatment information 34c over a predetermined period from each monitoring device 1 via the Internet line 2 (#31 in FIG. 3).
  • the probability of developing retinopathy of prematurity decreases to about 10% when the gestational age reaches 27 weeks.
  • approximately 40% of premature infants with a birth weight of less than 1000g are indicated for treatment, and premature infants with a birth weight of less than 1000g deteriorate rapidly (Retinopathy of prematurity). high risk of developing APROP).
  • the trained model generation device 3 extracts data on preterm infants with a gestational age of less than 28 weeks as the learning preterm infant information 34a and treatment information 34c (#32 in FIG. 3, filtering).
  • the trained model generation device 3 extracts, as the learning preterm infant information 34a and treatment information 34c, data relating to preterm infants whose total gestational age and treatment weeks are 40 weeks or less (# in FIG. 3). 32, filtering).
  • the trained model generation device 3 may extract, as the learning premature infant information 34a and the treatment information 34c, data on premature infants whose total gestational age and number of weeks at the time of treatment is 29 weeks or more and 40 weeks or less. .
  • the trained model generating device 3 excludes the premature infant information for learning 34a and the treatment information 34c, which are peculiar cases of early treatment determined by the doctor (#32 in FIG. 3, filtering).
  • the learning premature infant information 34a in this embodiment is time-series data of gestational age, weight, height, respiratory rate, heart rate, and arterial blood oxygen saturation for 206 preterm infants who underwent treatment at A hospital. , Apgar score at 5 minutes after birth, Apgar score at 1 minute after birth, sex, birth pattern, count, and onset (the terms are defined above).
  • the model generation unit 32 of the trained model generation device 3 receives the premature infant information for learning 34a as the input data (such as time-series data after birth regarding the weight, height and vital signs of the premature infant), and the teacher data as treatment information. 34c (without treatment, Type 1 ROP with treatment, APROP with treatment) and perform reinforcement learning to generate a learned model 10 (#33 to #36 in FIG. 3).
  • the learning feature amount calculation unit 33 calculates a plurality of learning feature amounts 34b from the learning premature infant information 34a (#34 in FIG. 3). , feature quantity calculation step).
  • This learning feature value 34b includes the number of days of gestation, daily average weight, weight difference, weight_SD, weight_SD difference, daily average height, height difference, height_SD, height_ SD difference, daily mean respiratory rate, respiratory rate difference, daily mean heart rate, heart rate difference, daily mean arterial oxygen saturation, arterial blood oxygen saturation difference, 5 minutes after birth postnatal Apgar score, 1 minute postnatal Apgar score, sex, birth morphology, count, onset (defined above). As shown in FIG. 8, when a plurality of learning feature values 34b calculated using the learning premature infant information 34a of 206 premature infants treated at A hospital are input to the learned model 10 and analyzed, , the degree of influence of each feature amount on the treatment information 34c is obtained.
  • the model generation unit 32 performs reinforcement learning using the input data as the learning feature amount 34b and the teacher data as the treatment information 34c (no treatment, type 1 ROP with treatment, APROP with treatment), and generates the learned model 10 ( #36 in FIG. 3).
  • the model generation unit 32 when the model generation unit 32 does not perform machine learning (#33 No in FIG. 3), it performs deep learning including a convolutional neural network (#35 in FIG. 3). In this deep learning, the model generating unit 32 uses premature infant information 34a for learning as input data (postnatal time-series data on the weight, height and vital signs of premature infants, etc.) and teacher data as treatment information 34c (without treatment). , Type 1 ROP with treatment, APROP with treatment) to generate a learned model 10 (#36 in FIG. 3).
  • the screening device 4 outputs the daily score (time-series AUC calculated from the ROC curve) of the trained model 10 to which the predictive premature infant information 46a is input at 20 days after birth (Fig. 7 See), the risk determination unit 43 determines the risk of progression of retinopathy of prematurity based on the score of Type 1 ROP or APROP (#37 in FIG. 3, risk determination step).
  • the predictive premature infant information 46a includes time-series data of gestational age, weight, height, respiratory rate, heart rate, and arterial blood oxygen saturation, Apgar score 5 minutes after birth, Apgar score 1 minute after birth, sex, and birth pattern. , counts (the terms are defined above). This score is calculated for each of no treatment, Type 1 ROP with treatment, and APROP with treatment.
  • the predictive preterm infant information 46a determined to have progression risk is used as onset data (Onset), and the treatment determination unit 44 learns the predictive preterm infant information 46a, which is the onset data. Then, based on the output values of the trained model 10, it is determined whether or not treatment for retinopathy of prematurity is indicated after 20 days of birth (#39 in FIG. 3, treatment determination step).
  • the trained model 10 in the present embodiment can distinguish between a spontaneously cured case (spontaneous regression) and a treatment adaptation case (disease progression) among the predictive premature infant information 46a determined to have a progression risk. can.
  • the treatment determination unit 44 includes the number of gestational days, daily average weight, weight difference, weight_SD, weight_SD difference, daily average height, height difference, height_SD, height_SD difference, daily average respiratory rate, respiratory rate difference, daily average heart rate, heart rate difference, daily average arterial blood oxygen saturation, Prediction feature values 46b composed of difference in arterial blood oxygen saturation, Apgar score 5 minutes after birth, Apgar score 1 minute after birth, sex, birth pattern, and count are input to the trained model 10 .
  • the treatment determination unit 44 performs postnatal treatment based on the output value of the learned model 10. After 20 days, it is determined whether or not the treatment for retinopathy of prematurity is indicated. More specifically, the treatment determination unit 44 determines that if the value of the treatment with treatment among the scores for no treatment, type 1 ROP with treatment, and APROP with treatment is the highest, the treatment determination unit 44 will predict that retinal photocoagulation will occur in the near future. (#40 Yes in FIG. 3), and the notification unit 45 notifies by a predetermined means (#41 in FIG. 3). In the example on the left side of FIG. 7, the score (AUC) of APROP with treatment became the highest at about 3 weeks after birth, so it is determined that the treatment is indicated.
  • FIG. 9 shows the progress prediction performance of retinopathy of prematurity using the trained model 10 that has undergone machine learning using the above-described learning feature value 34b as input data at Hospital A (206 premature infants).
  • FIG. 10 shows the progress prediction performance of retinopathy of prematurity using a trained model 10 that has undergone deep learning using the above-described learning premature infant information 34a as input data at Hospital A (206 premature infants).
  • the verification data shown in FIG. 9 is obtained by inputting the above-described prediction feature value 46b into the trained model 10, and evaluating the progress prediction performance of retinopathy of prematurity as a score ( Time-series AUC calculated from the ROC curve).
  • FIG. 10 is obtained by inputting the above-described predictive premature infant information 46a into the learned model 10, and evaluating the performance of predicting the progression of retinopathy of prematurity with scores ( Time-series AUC calculated from the ROC curve).
  • the upper part of FIG. 9 is 20 days after birth when the prediction feature value 46b of hospital A (206 premature babies) is input to the trained model 10 machine-learned with the learning feature value 34b of hospital A (206 premature babies).
  • the progression prediction performance (ROC curve) of the eye is shown, and the lower part of FIG. 20 shows the progress prediction performance (ROC curve) on the 20th day after birth when the prediction feature value 46b is input.
  • ROC curve progress prediction performance
  • the area under the ROC curve (AUC) without treatment in A hospital is 0.69
  • the AUC of APROP with treatment is 0.82
  • the AUC of Type 1 ROP with treatment is 0.69.
  • the area under the ROC curve (AUC) without treatment at B hospital is 0.66
  • the AUC of APROP with treatment is 0.83
  • the AUC of Type 1 ROP with treatment is It was 0.58, which was almost the same as the progression prediction performance in A hospital.
  • the trained model 10 machine-learned with the learning feature quantity 34b of the A hospital is a highly versatile model capable of predicting the progression of retinopathy of prematurity in the B hospital.
  • FIG. 10 shows the progress prediction performance when the premature infant information 46a for prediction of hospital B (59 preterm infants) is input to the trained model 10 that has undergone deep learning with the premature infant information 34a for learning of hospital A (206 preterm infants).
  • time-series data of AUC which is the time-series prediction performance calculated backward from the date of treatment or discharge.
  • the AUC is 0.8 or more at least 50 days before treatment, cases in which retinopathy of prematurity may progress to treatment indications can be determined in a timely manner.
  • the trained model 10 that has undergone deep learning with the learning premature infant information 34a of A hospital is a highly versatile model capable of predicting the progression of retinopathy of prematurity at B hospital.
  • the present embodiment can predict the progression of retinopathy of prematurity with higher accuracy than existing models (WINROP and CHOP-ROP models).
  • WINROP and CHOP-ROP models There are reports of fitting existing models in various countries, but we know that the accuracy varies significantly from country to country. It is speculated that this is due to differences in the medical standards of neonatal care.
  • the trained model 10 of this embodiment can also be used in facilities that manage different newborns. As a result, even those who do not have highly specialized knowledge and experience can judge treatment indications and start treatment at an appropriate time.
  • the trained model 10 in the present embodiment is highly versatile, capable of accurately predicting the progression of retinopathy of prematurity at an appropriate timing. In addition, if the total number of weeks of gestation and the number of weeks at the time of treatment is greater than 40 weeks, the risk of developing retinopathy of prematurity is extremely low.
  • the trained model 10 according to the present embodiment which is learned using the method, can accurately predict the progression of retinopathy of prematurity.
  • the two-step determination process provides a highly accurate screening method for retinopathy of prematurity.
  • the risk determination step in which the trained model 10 outputs the progression risk of retinopathy of prematurity may be omitted. Even in this case, treatment to determine whether or not retinopathy of prematurity is indicated for treatment after a predetermined number of days after birth using a trained model 10 that outputs scores for each of no treatment, Type 1 ROP with treatment, and APROP with treatment. The determination step can accurately predict the progression of retinopathy of prematurity.
  • the trained model 10 may be generated by machine learning other than a decision tree, or by deep learning other than a convolutional neural network. For example, a known learning method such as support vector machine or logistic regression can be used.
  • Premature infant information can include other parameters as long as they include postnatal chronological data on weight, height and vital signs.
  • the present disclosure can be used for a retinopathy of prematurity screening method, screening device, and trained model for predicting the progression of retinopathy of prematurity.
  • Screening device 10 Trained model 34a: Premature infant information for learning (premature infant information) 34b: feature quantity for learning (feature quantity) 46a: Preterm infant information for prediction (preterm infant information)

Landscapes

  • Health & Medical Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Public Health (AREA)
  • Medical Informatics (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Primary Health Care (AREA)
  • Epidemiology (AREA)
  • Pathology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Molecular Biology (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Surgery (AREA)
  • Animal Behavior & Ethology (AREA)
  • Veterinary Medicine (AREA)
  • Physics & Mathematics (AREA)
  • Biophysics (AREA)
  • Measuring And Recording Apparatus For Diagnosis (AREA)

Abstract

The purpose of the present invention is to provide: a method for screening for retinopathy of prematurity, which makes it possible to perform the prediction of the progression of retinopathy of prematurity at a proper timing and with high accuracy and which has broad utility; an apparatus for screening for retinopathy of prematurity; and a learned model. Provided is a method for screening for retinopathy of prematurity by performing the prediction of the progression of retinopathy of prematurity, the method including a treatment determination step of determining as to whether or not a treatment for retinopathy of prematurity can be applied to a premature baby on or after a predetermined number of days after birth on the basis of information about the premature baby, in which the premature baby is one whose fetus week number is smaller than a predetermined number of weeks, and the information includes time-series data after birth which are associated with the body weight, body height and vital signs of the premature baby.

Description

未熟児網膜症スクリーニング方法、スクリーニング装置及び学習済モデルScreening method for retinopathy of prematurity, screening device and learned model
　本開示は、医師による未熟児網膜症の診断を補助するために使用される、未熟児網膜症の進行予測を行う未熟児網膜症スクリーニング方法、スクリーニング装置及び学習済モデルに関する。 The present disclosure relates to a screening method for retinopathy of prematurity, a screening device, and a trained model for predicting the progression of retinopathy of prematurity, which are used to assist doctors in diagnosing retinopathy of prematurity.
　未熟児網膜症（Retinopathy Of Prematurity; ROP）は小児期の主要な失明原因疾患であり、世界で年間５万人が失明するとされ、今後更に増加すると予想されている。多くは発症しても自然軽快が見込める疾患であるが、特に未熟な超早産児では眼内出血や網膜剥離により失明に至ることがある。こうした重症例に対して網膜光凝固が標準治療として行われている。網膜光凝固は未熟児網膜症の進行抑制に有効であることが以前から示されているが、組織破壊を代償に失明を回避するものであり、予防的に行う治療方法ではない。最近では、光凝固と同等の効果を持つ治療として新たに血管内皮増殖因子(Vascular Endothelial Growth Factor)阻害薬（抗VEGF薬）の眼内投与（硝子体内投与）が実施されるようになっているが、全身発達への影響が懸念されており、やはり予防治療には適さない。一方、治療時期の遅れは、病勢の悪化により光凝固の効果が得られず、失明リスクを急激に上昇させることが明らかになっている。そこで、現在では国際分類に基づく病期判定と、米国の無作為試験（The Early Treatment for Retinopathy of Prematurity Study; ETROP）の結果に基づいた治療基準に従って、一定の重症度に達した症例に対して治療が実施されている。治療が奏功する治療時間域（Therapeutic time window）に対応するためには、頻回の眼底検査と迅速な治療が求められる。 Retinopathy of prematurity (ROP) is a major cause of blindness in childhood, and it is estimated that 50,000 people worldwide become blind each year, and the number is expected to increase further in the future. In many cases, spontaneous remission can be expected after the onset of the disease, but in immature, extremely premature infants, intraocular hemorrhage and retinal detachment may lead to blindness. Retinal photocoagulation is the standard treatment for such severe cases. Although retinal photocoagulation has been shown to be effective in suppressing the progression of retinopathy of prematurity, it avoids blindness at the expense of tissue destruction, and is not a preventative treatment. Recently, intraocular administration (intravitreal administration) of Vascular Endothelial Growth Factor inhibitors (anti-VEGF drugs) has become a new treatment that has the same effect as photocoagulation. However, there are concerns about the effects on systemic development, and it is not suitable for preventive treatment. On the other hand, it has been clarified that a delay in the timing of treatment results in an aggravation of the disease, making it impossible to obtain the effect of photocoagulation, resulting in a sharp increase in the risk of blindness. Therefore, currently, according to the stage determination based on the international classification and the treatment criteria based on the results of the US randomized trial (The Early Treatment for Retinopathy of Prematurity Study; ETROP), for cases that have reached a certain severity Treatment is being administered. Frequent fundus examinations and prompt treatment are required in order to respond to the therapeutic time window in which treatment is effective.
　未熟児網膜症スクリーニング方法として、例えば特許文献１に記載の技術が挙げられる。このスクリーニング方法は、被験者に由来する血液から、肥満細胞の脱顆粒により放出され得るトリプターゼをマーカー物質として検出し、未熟児網膜症の治療が必要か否かを判定する。 A method for screening for retinopathy of prematurity includes, for example, the technology described in Patent Document 1. In this screening method, tryptase, which can be released by degranulation of mast cells, is detected as a marker substance from the blood derived from the subject to determine whether treatment for retinopathy of prematurity is necessary.
　また、未熟児網膜症スクリーニング方法として、スウェーデンで開発されたモデル（WINROP）やアメリカから報告されたモデル（CHOP-ROP model）が知られている。WINROPは、在胎週数２３週以上３２週未満を対象にしており、在胎週数と出生体重及び出生後の体重を１週おきに入力していくと悪化の可能性がある症例にはアラームが表示される。CHOP-ROP modelは、WINROP同様に１週ごとの体重増加で評価し、Low risk群で診察回数を減らすことができる。 Also, as a screening method for retinopathy of prematurity, a model developed in Sweden (WINROP) and a model reported from the United States (CHOP-ROP model) are known. WINROP targets gestational age of 23 weeks or more and less than 32 weeks, and if the gestational age, birth weight, and postnatal weight are entered every other week, an alarm will be set for cases where there is a possibility of deterioration. Is displayed. The CHOP-ROP model, like WINROP, can be evaluated by weekly weight gain and can reduce the number of consultations in the low-risk group.
特開２０１４－２０８６０１号公報JP 2014-208601 A
　しかしながら、特許文献１に記載の方法では血液採取という侵襲的手段が必要となり現実的ではない。また、特許文献１に記載の方法により未熟児網膜症の治療が必要と判定された場合でも自然に治癒する可能性があるし、未熟児網膜症の治療が不要と判定された場合でも数日後には未熟児網膜症が発症して重症化する可能性もある。このため、未熟児網膜症の進行予測には、頻回の眼底検査が重要となる。 However, the method described in Patent Document 1 requires an invasive means of blood sampling, which is not realistic. In addition, even if it is determined that the treatment of retinopathy of prematurity is necessary by the method described in Patent Document 1, it may be cured spontaneously, and even if it is determined that the treatment of retinopathy of prematurity is unnecessary, it takes several days. Later, retinopathy of prematurity may develop and become severe. For this reason, frequent fundus examinations are important for predicting the progression of retinopathy of prematurity.
　スウェーデンやアメリカでは、非進行例のスクリーニング回数を減らす目的で、予測モデルが開発されているが、何れも出生後の体重増加を追跡する方法で、感度は高いが特異度は低いことが課題である。更に母集団の体重が異なる場合（途上国における早産児、重症化リスクの高い出生体重が1000g未満の早産児）には、精度は低いことが示されている。 In Sweden and the United States, predictive models have been developed with the aim of reducing the number of screenings for non-progressive cases. be. Furthermore, accuracy has been shown to be low when population weights differ (preterm infants in developing countries, preterm infants with birth weight <1000 g at high risk of severe disease).
　そこで、適切なタイミングで未熟児網膜症の進行予測を精度良く行うことが可能な、汎用性の高い未熟児網膜症スクリーニング方法、スクリーニング装置及び学習済モデルが望まれている。 Therefore, there is a demand for a highly versatile retinopathy screening method, screening device, and trained model that can accurately predict the progression of retinopathy of prematurity at an appropriate timing.
　本開示の一態様は、未熟児網膜症の進行予測を行う未熟児網膜症スクリーニング方法であって、在胎週数が所定週未満である早産児の体重、身長及びバイタルサインに関する出生後の時系列データを含む早産児情報に基づいて、所定の生後日数以降で未熟児網膜症の治療適応となるか否かを判定する治療判定工程を含む点にある。また、本開示の一態様は、未熟児網膜症の進行予測を行うスクリーニング装置であって、在胎週数が所定週未満である早産児の体重、身長及びバイタルサインに関する出生後の時系列データを含む早産児情報に基づいて、所定の生後日数以降で未熟児網膜症の治療適応となるか否かを判定する治療判定部を備えた点にある。 One aspect of the present disclosure is a screening method for retinopathy of prematurity for predicting the progression of retinopathy of prematurity, comprising a postnatal time series of weight, height, and vital signs of a premature infant whose gestational age is less than a predetermined week. The method includes a treatment determination step for determining whether or not treatment for retinopathy of prematurity is indicated after a predetermined number of days after birth, based on premature infant information including data. In addition, one aspect of the present disclosure is a screening device that predicts the progression of retinopathy of prematurity, and includes postnatal chronological data on the weight, height, and vital signs of a premature infant whose gestational age is less than a predetermined week. The present invention is characterized by including a treatment determination unit that determines whether or not treatment for retinopathy of prematurity is indicated after a predetermined number of days after birth, based on premature infant information.
　本方法又は本装置では、体重、身長及びバイタルサインに関する出生後の時系列データを用いて未熟児網膜症の治療適応となるか否かを判定することにより、未熟児網膜症の進行を予測することができる。未熟児網膜症の発症から治療適応となるまでの進行過程には多くの要素が複雑に関連し、その要素は経時的に変動する。本構成では、複数の時系列データを用いて未熟児網膜症の進行を高い精度で予測する方法又は装置となっている。 In this method or device, the progression of retinopathy of prematurity is predicted by determining whether treatment for retinopathy of prematurity is indicated using postnatal time-series data on weight, height, and vital signs. be able to. Many factors are intricately related to the progression of retinopathy of prematurity from onset to treatment indication, and these factors change over time. This configuration provides a method or apparatus for predicting the progression of retinopathy of prematurity with high accuracy using a plurality of time-series data.
　本方法又は本装置は、未熟性及び全身状態の経時変化を含む情報を用いる。例えば、未熟性は出生時の在胎週数等を指標とし、生後の全身状態は心拍数、呼吸及び血中酸素濃度等のバイタルサイン、更に体重、身長を含む時系列データを指標として用いる。また、所定の生後日数以降で未熟児網膜症の治療適応となるか否かを判定している。本方法又は本装置を適応できると見込まれる具体的な実施例として、新生児集中治療室において、早産児のバイタルサインを経時的にモニタリングする機器にプログラムを搭載し、近いうちに未熟児網膜症の治療適応となる症例に警告サインを示す診断補助機器が挙げられる。 This method or this device uses information including changes in immaturity and general condition over time. For example, the gestational age at birth is used as an index for immaturity, and the general condition after birth is used as an index for time-series data including vital signs such as heart rate, respiration and blood oxygen concentration, as well as body weight and height. In addition, it is determined whether or not treatment for retinopathy of prematurity is indicated after a predetermined number of days after birth. As a specific example to which the present method or apparatus is expected to be applicable, a program is installed in equipment for chronological monitoring of vital signs of premature infants in the neonatal intensive care unit, and in the near future, retinopathy of prematurity will occur. These include diagnostic aids that provide warning signs for treatment-indicated cases.
　本方法又は本装置によれば、既存のモデル（WINROPやCHOP-ROP model）に比べて高い精度で未熟児網膜症の進行を予測できる。未熟児網膜症に対して、時期を逸することなく適切な治療を行うことで、失明リスクは減少する。一方で、正確な病期判定と治療に習熟した眼科医の不足は、世界的に深刻な問題となっている。しかし、本方法又は本装置によれば、高度に専門的な知識と経験を有する者でなくとも治療適応を判断し、適切な時期に治療を開始することを可能にする。また、一般的には、眼科医による倒像鏡による眼底検査もしくは接触型眼底カメラにより得られた画像から未熟児網膜症の病期判定を行うが、何れの方法も脆弱な早産児にとって全身負担が大きく、診察中や画像撮影中に心拍低下などの発生リスクを伴う。本方法又は本装置では、眼底検査や画像撮影を必要とせず、早産児の体重、身長及びバイタルサインに関する出生後の時系列データを含む早産児情報に基づいて治療適応となるか否かを判定しているため、不要な診察を減らすことにも繋がる。つまり、医療への貢献だけでなく、視覚障害による医療費、生産性の低下又は社会的ケアに係るコストを大幅に減少させることに繋がり、社会経済にも資する。 According to this method or device, the progression of retinopathy of prematurity can be predicted with higher accuracy than existing models (WINROP and CHOP-ROP models). Timely and appropriate treatment of retinopathy of prematurity reduces the risk of blindness. On the other hand, the shortage of ophthalmologists skilled in accurate staging and treatment has become a serious problem worldwide. However, according to this method or this device, even a person who does not have highly specialized knowledge and experience can judge treatment indications and start treatment at an appropriate time. In general, the stage of retinopathy of prematurity is determined from images obtained by ophthalmologists using an indirect ocular fundus examination or a contact-type fundus camera. is large, and there is a risk of occurrence such as hypocardiac during examination and imaging. The method or device determines suitability for treatment based on preterm infant information, including postnatal chronological data on weight, height, and vital signs of preterm infants, without the need for fundus examination or imaging. This also helps reduce unnecessary medical examinations. In other words, it not only contributes to medical care, but also contributes to society and the economy by significantly reducing medical costs due to visual impairment, lower productivity, and costs related to social care.
　このように、適切なタイミングで未熟児網膜症の進行予測を精度良く行うことが可能な、汎用性の高い未熟児網膜症スクリーニング方法又はスクリーニング装置となっている。 In this way, it is a highly versatile screening method or screening device for retinopathy of prematurity that can accurately predict the progression of retinopathy of prematurity at an appropriate timing.
　本方法に関する他の態様は、前記所定の生後日数時点において算出されたType1 ROP又はAPROPのスコアに基づいて未熟児網膜症の進行リスクを判定するリスク判定工程を更に含み、前記リスク判定工程で前記進行リスクが有ると判定された前記早産児のみ前記治療判定工程を実行する点にある。また、本装置に関する他の態様は、前記所定の生後日数時点において算出されたType1 ROP又はAPROPのスコアに基づいて未熟児網膜症の進行リスクを判定するリスク判定部を更に備え、前記治療判定部は、前記リスク判定部で前記進行リスクが有ると判定された前記早産児のみを対象として前記治療適応となるか否かを判定する点にある。 Another embodiment of this method further comprises a risk determination step of determining the risk of progression of retinopathy of prematurity based on the Type 1 ROP or APROP score calculated at the predetermined number of days after birth, wherein the The point is that the treatment determination step is performed only for the premature infant determined to have a progression risk. Further, another aspect of the present device further comprises a risk determination unit that determines the risk of progression of retinopathy of prematurity based on the Type 1 ROP or APROP score calculated at the predetermined number of days after birth, and the treatment determination unit The point is to determine whether or not the treatment is indicated for only the premature infant determined by the risk determination unit to have the progression risk.
　本方法又は本装置では、未熟児網膜症の進行リスクを判定するため、未熟児網膜症の潜在的な進行度合いを推定することができる。そして、進行リスクが有ると判定された早産児のみ治療適応となるか否かを識別するため、２段階の判定工程により精度の高い未熟児網膜症スクリーニング方法又はスクリーニング装置となっている。 With this method or device, the latent degree of progression of retinopathy of prematurity can be estimated in order to determine the risk of progression of retinopathy of prematurity. In order to discriminate whether or not only premature infants determined to have a risk of progression are eligible for treatment, the two-step determination process provides a highly accurate screening method or screening device for retinopathy of prematurity.
　本方法に関する他の態様は、前記バイタルサインは、前記早産児の心拍数、呼吸数及び動脈血酸素飽和度の少なくとも１つである点にある。 Another aspect of the method is that the vital sign is at least one of heart rate, respiratory rate and arterial blood oxygen saturation of the premature infant.
　このようなバイタルサインであれば、新生児集中治療室に設置された既存のモニタリング機器で取得できるため、新たに装置を開発する必要なく、効率的である。 Such vital signs can be acquired with the existing monitoring equipment installed in the neonatal intensive care unit, so there is no need to develop new equipment, which is efficient.
　本方法に関する他の態様は、前記早産児情報は、前記早産児の在胎日数及びアプガースコアの少なくとも１つを含んでいる点にある。 Another aspect of this method is that the preterm infant information includes at least one of the gestational age and Apgar score of the preterm infant.
　このように、早産児情報として、生後の全身状態に加えて、早産児の在胎日数及びアプガースコアを用いれば、未熟児網膜症の進行予測を更に精度良く行うことができる。 In this way, the progression of retinopathy of prematurity can be predicted more accurately by using the gestational age and Apgar score of premature infants in addition to the postnatal general condition as premature infant information.
　本開示の一態様は、コンピュータにより機能する学習済モデルであって、ツリー構造に並んだ複数の分岐点からなる決定木で構成され、未熟児網膜症の治療を行った在胎週数が所定週未満である早産児の体重、身長及びバイタルサインに関する出生後の時系列データを含む早産児情報に基づいて演算した特徴量が入力され、夫々の前記分岐点での評価値を合算することにより、未熟児網膜症の治療要否のスコアを出力する点にある。 One aspect of the present disclosure is a trained model that functions by a computer, is composed of a decision tree consisting of a plurality of branch points arranged in a tree structure, and has a predetermined number of weeks of gestation when treatment for retinopathy of prematurity is performed. A feature value calculated based on preterm infant information including postnatal time-series data on the weight, height and vital signs of a preterm infant that is less than is input, and by summing the evaluation values at each of the branch points, The point is to output a score indicating the necessity of treatment for retinopathy of prematurity.
　本構成のように決定木による機械学習をした学習済モデルであれば、適切なタイミングで未熟児網膜症の進行予測を精度良く行うことが可能な、汎用性の高いものとなる。また、学習済モデルに入力するために、早産児情報を加工して特徴量を演算すれば、未熟児網膜症の進行予測を更に精度良く行うことができる。 If it is a trained model that has undergone machine learning using a decision tree like this configuration, it will be highly versatile, making it possible to accurately predict the progression of retinopathy of prematurity at the appropriate timing. Further, by processing the premature infant information and calculating the feature amount for inputting to the trained model, it is possible to predict the progress of retinopathy of prematurity with higher accuracy.
　本開示の一態様は、コンピュータにより機能する学習済モデルであって、畳み込みニューラルネットワークを含む深層学習で生成され、未熟児網膜症の治療を行った在胎週数が所定週未満である早産児の体重、身長及びバイタルサインに関する出生後の時系列データを含む早産児情報が入力され、未熟児網膜症の治療要否のスコアを出力する点にある。 One aspect of the present disclosure is a computer-functioning trained model, generated by deep learning including a convolutional neural network, for premature infants treated for retinopathy of prematurity with a gestational age of less than a predetermined number of weeks. Premature infant information including postnatal chronological data on weight, height and vital signs is input, and a score for the need for treatment of retinopathy of prematurity is output.
　本構成のように畳み込みニューラルネットワークを含む深層学習をした学習済モデルであれば、適切なタイミングで未熟児網膜症の進行予測を精度良く行うことが可能な、汎用性の高いものとなる。 A trained model that has undergone deep learning, including a convolutional neural network, as in this configuration, is highly versatile and can accurately predict the progression of retinopathy of prematurity at an appropriate timing.
　本モデルに関する他の態様は、前記早産児情報は、前記早産児の在胎週数及び治療時週数の合計が４０週以下の前記早産児から得た情報である点にある。 Another aspect of this model is that the premature infant information is information obtained from the premature infant whose total gestational age and number of weeks at the time of treatment is 40 weeks or less.
　在胎週数及び治療時週数の合計が４０週より大きければ、未熟児網膜症の進行リスクが極めて小さいため、本構成における学習済モデルは、未熟児網膜症の進行予測を精度良く行うことができる。 If the total number of weeks of gestation and number of weeks at the time of treatment is greater than 40 weeks, the risk of progression of retinopathy of prematurity is extremely low. can.
　本モデルに関する他の態様は、前記早産児情報は、医師判断により早期治療した前記早産児から得た情報を除外したものである点にある。 Another aspect of this model is that the premature infant information excludes information obtained from the premature infants who were treated early based on the doctor's judgment.
　このように、医師判断により早期に治療した特異例を除外すれば、良好な学習済モデルを提供できる。 In this way, a good trained model can be provided by excluding anomalous cases that were treated early at the doctor's discretion.
　本モデルに関する他の態様は、前記バイタルサインは、前記早産児の心拍数、呼吸数及び動脈血酸素飽和度の少なくとも１つである点にある。 Another aspect of this model is that the vital sign is at least one of heart rate, respiratory rate and arterial blood oxygen saturation of the premature infant.
　このようなバイタルサインであれば、新生児集中治療室に設置された既存のモニタリング機器で取得されているため、学習済モデルを構築するための入力データを多く確保できる。 Such vital signs are acquired by existing monitoring equipment installed in the neonatal intensive care unit, so it is possible to secure a large amount of input data for building a trained model.
　本モデルに関する他の態様は、前記早産児情報は、前記早産児の在胎日数及びアプガースコアの少なくとも１つを含んでいる点にある。 Another aspect of this model is that the preterm infant information includes at least one of the gestational age and Apgar score of the preterm infant.
　このように生後の全身状態に加えて、早産児の在胎日数及びアプガースコアを用いて学習済モデルを構築すれば、未熟児網膜症の進行予測を更に精度良く行うことができる。 In this way, if a trained model is constructed using the gestational age and Apgar score of premature infants in addition to the general condition after birth, the progression of retinopathy of prematurity can be predicted with even higher accuracy.
は、本実施形態に係るスクリーニング方法を実現するシステム全体図である。1] is an overall view of a system for realizing a screening method according to this embodiment. [FIG. は、本実施形態に係るスクリーニング装置のブロック図である。1 is a block diagram of a screening device according to this embodiment; FIG. は、本実施形態に係るスクリーニング方法を実現するフロー図である。1] is a flowchart for realizing a screening method according to the present embodiment. [FIG. は、本実施形態に係るスクリーニング方法の説明図である。1] is an explanatory diagram of a screening method according to the present embodiment. [FIG. は、在胎日数と治療実績の関係を示す図である。[ Fig. 3] Fig. 3 is a diagram showing the relationship between gestational age and treatment results. は、治療時週数又は在胎週数及び治療時週数の合計と治療実績の関係を示す図である。FIG. 10 is a diagram showing the relationship between the number of weeks at the time of treatment or the gestational age and the total number of weeks at the time of treatment and the treatment results. は、リスク判定工程の一例を示す図である。is a diagram showing an example of a risk determination process. は、機械学習における特量量ごとの寄与度を示す図である。4 is a diagram showing the degree of contribution of each feature amount in machine learning; FIG. は、機械学習した学習済モデルを用いて治療判定工程を実行した一例のＲＯＣ曲線図である。FIG. 10 is an example ROC curve diagram in which a treatment decision process is performed using a machine-learned trained model; は、深層学習した学習済モデルを用いて治療判定工程を実行した一例のＡＵＣ図である。FIG. 4 is an AUC diagram of an example of performing a treatment determination process using a trained model that has undergone deep learning;
　以下に、本開示に係る未熟児網膜症スクリーニング方法、スクリーニング装置及び学習済モデルの実施形態について、図面に基づいて説明する。ただし、以下の実施形態に限定されることなく、その要旨を逸脱しない範囲内で種々の変形が可能である。 Embodiments of the retinopathy of prematurity screening method, screening device, and trained model according to the present disclosure will be described below based on the drawings. However, without being limited to the following embodiments, various modifications are possible without departing from the scope of the invention.
　図１，図２を用いて、未熟児網膜症スクリーニング方法に用いられるシステム構成について説明する。 The system configuration used for the screening method for retinopathy of prematurity will be explained using Figures 1 and 2.
　早産児の体重、身長及びバイタルサインに関する出生後の時系列データを含む早産児情報を取得するための１又は複数のモニタリング機器１は、インターネット回線２に接続される。また、インターネット回線２には、学習済モデル生成装置３と、スクリーニング装置４と、ＡＩ９（人工知能）とが接続される。ここで、ＡＩ９は、インターネット回線２上に設けられても良いし、学習済モデル生成装置３内に設けられても良い。また、学習済モデル生成装置３とスクリーニング装置４とは同一の装置であっても良いし、スクリーニング装置４をモニタリング機器１に内蔵しても良く、夫々の機能を単独又は併合して用いることができる。また、スクリーニング装置４は、保育器に備え付けられたモニタ装置や新生児集中治療室内の専用機器としても良く、未熟児網膜症の進行予測を行う様々な診断補助機器として活用可能である。 　One or more monitoring devices 1 for obtaining preterm infant information, including postnatal chronological data on the weight, height and vital signs of preterm infants, are connected to the Internet line 2. Also, the Internet line 2 is connected with a trained model generation device 3, a screening device 4, and an AI 9 (artificial intelligence). Here, the AI 9 may be provided on the Internet line 2 or may be provided in the trained model generation device 3 . In addition, the trained model generation device 3 and the screening device 4 may be the same device, or the screening device 4 may be incorporated in the monitoring device 1, and their functions may be used singly or in combination. can. Further, the screening device 4 may be a monitor device installed in an incubator or a dedicated device in a neonatal intensive care unit, and can be used as various diagnostic auxiliary devices for predicting the progression of retinopathy of prematurity.
　モニタリング機器１は、新生児集中治療室において早産児のバイタルサインを経時的にモニタリングする機器や早産児の体重，身長を周期的に測定する機器である。本実施形態の早産児情報は、在胎週数が所定週（例えば２８週）未満である早産児の体重、身長及びバイタルサインを含んでいる。この所定週は、３６週以下（所謂、早産児）、好ましくは３２週以下、更に好ましくは２８週以下（所謂、超早産児）である（以下、同様）。本実施形態では、未熟性の指標として出生時の在胎週数を用いているが、出生体重も指標として併用しても良い。この出生体重は、２５００ｇ未満（所謂、低出生体重児）、好ましくは１５００ｇ未満（所謂、極低出生体重児）、更に好ましくは１０００ｇ未満（所謂、超低出生体重児）である。バイタルサインは、早産児の心拍数、呼吸数及び動脈血酸素飽和度の少なくとも１つであり、少なくとも１分おきに取得される。このバイタルサインは、早産児の生命情報であれば、血圧等を含んでいても良い。早産児情報は、週に３回の頻度で取得された早産児の体重、週に１回の頻度で取得された早産児の身長、早産児の在胎日数、及び生後１分後と５分後のアプガースコアの少なくとも１つを含んでいることが好ましい。早産児情報の取得頻度は、バイタルサインについて１秒おきや１０分おき、体重や身長について毎日に設定する等、特に限定されない。アプガースコアは、出生直後の新生児の状態を評価するもので、皮膚色、心拍数、反応、筋緊張、呼吸の５項目をトータル１０点満点で評価したものである。 The monitoring device 1 is a device that monitors the vital signs of preterm infants over time and a device that periodically measures the weight and height of preterm infants in the neonatal intensive care unit. The preterm infant information of this embodiment includes the weight, height, and vital signs of a preterm infant whose gestational age is less than a predetermined number of weeks (for example, 28 weeks). This predetermined week is 36 weeks or less (so-called premature infants), preferably 32 weeks or less, more preferably 28 weeks or less (so-called very premature infants) (hereinafter the same). Although the gestational age at birth is used as an index of immaturity in this embodiment, birth weight may also be used as an index. This birth weight is less than 2500 g (so-called low birth weight infants), preferably less than 1500 g (so-called very low birth weight infants), more preferably less than 1000 g (so-called very low birth weight infants). Vital signs are at least one of heart rate, respiratory rate and arterial oxygen saturation of premature infants and are obtained at least every minute. The vital signs may include blood pressure and the like as long as they are vital information of the premature infant. Premature infant information includes weight of preterm infants obtained three times a week, height of preterm infants obtained once a week, gestational age of preterm infants, and 1 and 5 minutes after birth. It preferably contains at least one of the later Apgar scores. The acquisition frequency of premature infant information is not particularly limited, and may be set every second or every 10 minutes for vital signs, or every day for weight and height. The Apgar score evaluates the condition of a newborn immediately after birth, and evaluates the five items of skin color, heart rate, reaction, muscle tone, and respiration on a total of 10 points.
　図２に示すように、学習済モデル生成装置３は、第一通信部３１、モデル生成部３２、学習用特徴量演算部３３及び第一記憶部３４を備えている。 As shown in FIG. 2, the trained model generation device 3 includes a first communication unit 31, a model generation unit 32, a learning feature value calculation unit 33, and a first storage unit .
　第一通信部３１は、インターネット回線２を介して、モニタリング機器１、スクリーニング装置４又はＡＩ９等との間でデータの送受信を行うインターフェースである。第一通信部３１は、モニタリング機器１から直接データを受信しても良いし、モニタリング機器１で取得されたデータをサーバ（不図示）に蓄積し、このサーバから蓄積されたデータを受信しても良い。 The first communication unit 31 is an interface that transmits and receives data to and from the monitoring device 1, the screening device 4, the AI 9, etc. via the Internet line 2. The first communication unit 31 may receive data directly from the monitoring device 1, or accumulate data acquired by the monitoring device 1 in a server (not shown) and receive the accumulated data from this server. Also good.
　第一記憶部３４は、ＨＤＤやＳＳＤ等の一時的でない記憶媒体又はＲＡＭ等の一時的な記憶媒体で構成されており、プロセッサにより実行されるプログラムやアプリケーションを記憶している。この第一記憶部３４は、第一通信部３１を介して取得したモニタリング機器１の学習用早産児情報３４ａを記憶している。学習用早産児情報３４ａは、在胎週数が所定週（例えば２８週）未満である早産児の体重、身長及びバイタルサインに関する出生後の時系列データを含んでいる。また、学習用早産児情報３４ａは、早産児の在胎日数及びアプガースコア等を含んでいる。この学習用早産児情報３４ａは、早産児の在胎週数及び治療時週数の合計が４０週以下の早産児から得た情報であることが好ましい。また、学習用早産児情報３４ａは、医師判断により早期治療した早産児から得た情報を除外したものあることが好ましい。 The first storage unit 34 is composed of a non-temporary storage medium such as an HDD or SSD or a temporary storage medium such as a RAM, and stores programs and applications to be executed by the processor. The first storage unit 34 stores learning premature infant information 34 a of the monitoring device 1 acquired via the first communication unit 31 . The learning preterm infant information 34a includes postnatal chronological data on the weight, height, and vital signs of preterm infants whose gestational age is less than a predetermined week (for example, 28 weeks). In addition, the premature infant information for learning 34a includes the gestational age and Apgar score of the premature infant. This premature infant information for learning 34a is preferably information obtained from a premature infant whose total gestational age and number of weeks at the time of treatment is 40 weeks or less. In addition, it is preferable that the premature infant information for learning 34a excludes information obtained from premature infants who were treated early by doctor's judgment.
　第一記憶部３４は、学習用早産児情報３４ａに関連付けられた治療情報３４ｃを記憶している。この治療情報３４ｃは、治療無し、治療有りType1 ROP、治療有りAPROP（Aggressive Posterior Retinopathy Of Prematurity）に区分される。未熟児網膜症の国際分類Classic ROP（典型ROP）にはType1 ROPとType2 ROPがあり、Type1が治療適応となっており、Type2が治療適応未満（治療無し）となっている。一方、Classic ROPとは別に急速に悪化するタイプをAPROPと呼び、こちらも治療適応となっている。Type1 ROP（国際分類）は、1型3期（厚生省分類）に相当し、APROP（国際分類）は2型（劇症型）に相当する。以下、治療適応とは、Type1 ROPと診断されてから７２時間以内に治療を実施すること、又は、APROPの初期兆候があれば迅速に治療を実施することを言う（以下、同様）。治療適応があるROPは、Type1 ROPのうち、plus diseaseを伴うzone1 ROP、plus diseaseを伴わないzone1 stage3 ROP、及びplus diseaseを伴うzone2 stage3 ROPの何れか、又は、APROPである。ここで、plus diseaseとは、網膜血管の拡張や蛇行がみられるものであり、zone1とは、視神経乳頭を中心として乳頭-黄斑距離の２倍を半径とする円内の領域であり、zone2とは、乳頭から鼻側鋸状縁までを半径とする円内の領域であり、stage3とは、網膜外線維血管増殖のことである。 The first storage unit 34 stores treatment information 34c associated with learning premature infant information 34a. This treatment information 34c is classified into no treatment, Type 1 ROP with treatment, and APROP (Aggressive Posterior Retinopathy Of Prematurity) with treatment. The International Classification of Classic ROP (typical ROP) for retinopathy of prematurity includes Type 1 ROP and Type 2 ROP, with Type 1 being indicated for treatment and Type 2 being less than indicated for treatment (no treatment). On the other hand, a type that rapidly worsens apart from Classic ROP is called APROP, which is also indicated for treatment. Type 1 ROP (international classification) corresponds to type 1 stage 3 (Ministry of Health and Welfare classification), and APROP (international classification) corresponds to type 2 (fulminant type). In the following, “treatment indication” means to implement treatment within 72 hours after being diagnosed with Type 1 ROP, or to implement treatment promptly if there is an early sign of APROP (hereinafter the same). Type 1 ROP with indications for treatment is either zone1 ROP with plus disease, zone1 stage3 ROP without plus disease, or zone2 stage3 ROP with plus disease, or APROP. Here, plus disease is dilation or tortuousness of retinal blood vessels, zone 1 is the area within a circle centered on the optic nerve head with a radius twice the distance between the head and the macula, and zone 2 is the area within a circle with a radius from the papilla to the nasal serrated margin, and stage 3 is extraretinal fibrovascular proliferation.
　第一記憶部３４は、学習済モデル１０を記憶している。学習済モデル１０は、コンピュータにより機能するモデルであって、機械学習又は深層学習の教師データ有り学習により得られる。更に、第一記憶部３４は、機械学習するために、学習用特徴量演算部３３により演算される学習用特徴量３４ｂを記憶している。 The first storage unit 34 stores the trained model 10. The trained model 10 is a model that functions by a computer and is obtained by machine learning or deep learning with supervised data. Further, the first storage unit 34 stores a learning feature quantity 34b calculated by the learning feature quantity calculation unit 33 for machine learning.
　機械学習は、ツリー構造に並んだ複数の分岐点からなる決定木で構成されている。この機械学習は、決定木の各分岐点で特徴量の評価を行い、分岐点毎に、評価結果に応じた評価値が付与されていく構造である。そして、決定木の分岐に沿って評価値が合算されて未熟児網膜症の進行予測情報が求められる。機械学習は、複数の決定木が関連して設けられる、ＸＧＢｏｏｓｔやＲａｎｄｏｍ　Ｆｏｒｅｓｔ、ＬｉｇｈｔＧＢＭ、ＣａｔＢｏｏｓｔ、ＡｄａＢｏｏｓｔ等のアンサンブルモデルを用いて行っても良い。 Machine learning consists of a decision tree consisting of multiple branch points arranged in a tree structure. This machine learning has a structure in which a feature amount is evaluated at each branch point of a decision tree, and an evaluation value corresponding to the evaluation result is given to each branch point. Then, the evaluation values are summed up along the branches of the decision tree to obtain progression prediction information for retinopathy of prematurity. Machine learning may be performed using an ensemble model such as XGBoost, Random Forest, LightGBM, CatBoost, AdaBoost, etc., in which a plurality of decision trees are associated.
　深層学習は、公知の畳み込みニューラルネットワーク（ＣＮＮ，ＤＣＧＡＮ等）を含むＡＩ９によって実行される。畳み込みニューラルネットワークとは、人間の神経回路を模した深い階層のモデルを構築し、未熟児網膜症の進行予測を推論する。この深層学習は、インターネット回線２を介して提供されている公知のアプリケーションで構成されている。 Deep learning is performed by AI9, including well-known convolutional neural networks (CNN, DCGAN, etc.). A convolutional neural network constructs a deep-hierarchical model imitating a human neural circuit and infers the progression prediction of retinopathy of prematurity. This deep learning consists of known applications provided via the Internet line 2 .
　モデル生成部３２は、プロセッサを備えており、学習済モデル１０を生成する。プロセッサは、ＡＳＩＣ，ＦＧＰＡ，ＣＰＵ又は第一記憶部３４に記憶されたアプリケーション等を実行するための他のハードウェアを含んでいる（以下、同様）。機械学習により学習済モデル１０を生成する場合、モデル生成部３２は、入力データが学習用特徴量３４ｂで、教師データが治療情報３４ｃ（治療無し、治療有りType1 ROP、治療有りAPROP）として強化学習させ、学習済モデル１０を生成する。学習用特徴量３４ｂは、学習用早産児情報３４ａ（早産児の体重、身長及びバイタルサインに関する出生後の時系列データ、早産児の在胎日数及びアプガースコア等）を加工して演算されるが、詳細は後述する。一方、深層学習により学習済モデル１０を生成する場合、モデル生成部３２は、入力データが学習用早産児情報３４ａ（早産児の体重、身長及びバイタルサインに関する出生後の時系列データ等）で、教師データが治療情報３４ｃ（治療無し、治療有りType1 ROP、治療有りAPROP）として強化学習させ、学習済モデル１０を生成する。 The model generation unit 32 includes a processor and generates the trained model 10. The processor includes ASIC, FGPA, CPU, or other hardware for executing applications or the like stored in the first storage unit 34 (the same applies hereinafter). When generating the learned model 10 by machine learning, the model generation unit 32 performs reinforcement learning with the input data as the learning feature amount 34b and the teacher data as the treatment information 34c (no treatment, Type 1 ROP with treatment, APROP with treatment). to generate a trained model 10. The feature value for learning 34b is calculated by processing the premature infant information for learning 34a (time-series data after birth regarding the weight, height and vital signs of the premature infant, the gestational age and Apgar score of the premature infant, etc.). Details will be described later. On the other hand, when generating the trained model 10 by deep learning, the model generating unit 32 uses the input data as premature infant information 34a for learning (time-series data after birth regarding the weight, height and vital signs of the premature infant, etc.), Reinforcement learning is performed with teacher data as treatment information 34c (no treatment, Type 1 ROP with treatment, APROP with treatment), and a learned model 10 is generated.
　学習用特徴量演算部３３は、プロセッサを備えており、学習用早産児情報３４ａから複数の学習用特徴量３４ｂを演算し、演算された複数の学習用特徴量３４ｂを学習済モデル１０に入力して解析し、各特徴量の治療情報３４ｃへの影響度を求める。図８には、複数の学習用特徴量３４ｂの治療情報３４ｃへの影響度が例示されている。影響度の高い学習用特徴量３４ｂの順に、体重、身長、身長_ＳＤ、在胎日数、生後５分後のアプガースコア（Apgar Score; 5min）、体重_ＳＤ、生後１分後のアプガースコア（Apgar Score; 1min）、動脈血酸素飽和度（SpO2 %）、心拍数（HR bpm）、性別（M or F）、発症、出生形態（単胎、双胎、品胎）、カウント（count）、呼吸数（RESP/min）、呼吸数の差分（RESP/min.delta）、心拍数の差分（HR bpm.delta）、体重_ＳＤの差分、体重の差分、動脈血酸素飽和度の差分（SpO2 %.delta）、身長_ＳＤの差分、身長の差分となっている。なお、学習用特徴量演算部３３は、影響度の高い順に複数（例えば１０個）の指標を抽出し、モデル生成部３２は、この抽出された複数の指標を学習用特徴量３４ｂとして用いても良い。 The learning feature quantity calculation unit 33 includes a processor, calculates a plurality of learning feature quantities 34b from the learning premature infant information 34a, and inputs the calculated plurality of learning feature quantities 34b to the trained model 10. Then, the degree of influence of each feature amount on the treatment information 34c is obtained. FIG. 8 illustrates the degree of influence of a plurality of learning feature values 34b on the treatment information 34c. Weight, height, height_SD, gestational age, Apgar score 5 minutes after birth (Apgar score; 5min), weight_SD, Apgar score 1 minute after birth (Apgar score 1 min), arterial oxygen saturation (SpO2 %), heart rate (HR bpm), sex (M or F), onset, birth type (single, twin, conception), count, respiratory rate ( RESP/min), respiration rate difference (RESP/min.delta), heart rate difference (HR bpm.delta), weight_SD difference, body weight difference, arterial blood oxygen saturation difference (SpO2 %.delta) , height_SD difference, and height difference. Note that the learning feature quantity calculation unit 33 extracts a plurality of (for example, 10) indices in descending order of influence, and the model generation unit 32 uses the extracted plurality of indices as the learning feature quantity 34b. Also good.
　体重とは、週に３回取得した出生後の早産児の体重を１日平均値（又は１時間平均値）として前方補間したものである。身長とは、週に１回取得した出生後の早産児の身長を１日平均値（又は１時間平均値）として前方補間したものである。_ＳＤとは、平均値からのばらつきの大きさ、つまり分布の幅をＳＤ（標準偏差）という数値で表したものである。動脈血酸素飽和度、心拍数及び呼吸数は、出生後の早産児のバイタルサインを１分おきに取得したデータからゼロ値を除去して、１日平均値（又は１時間平均値）として補間したものである。発症とは、出生後所定の回数実行される医師による未熟児網膜症の所見（未熟児網膜症が発症したか否か）である。カウントとは、出生日からの経過日数を単位時間に変換したものである。差分とは、夫々のパラメータの測定ごとの差分を算出して、差分特徴量としたものである。なお、本実施形態における入力データである体重、身長及びバイタルサインは１日平均値を用いているが、１分平均値～２日平均値であれば良く、好ましくは１時間平均値～１日平均値、更に好ましくは１時間平均値又は１日平均値である（以下、同様）。入力データとして、２日超の平均値を用いた場合は予測精度が悪くなり、１分未満の平均値を用いた場合はデータ量が多く、演算速度の低下やノイズの混入を招くおそれがある。 Weight is forward-interpolated from the weight of postnatal premature infants obtained three times a week as a daily average (or hourly average). The height is obtained by forward-interpolating the height of the preterm infant after birth obtained once a week as a daily average value (or an hourly average value). _SD is a numerical value called SD (standard deviation) that represents the degree of variation from the average value, that is, the width of the distribution. Arterial oxygen saturation, heart rate, and respiratory rate were interpolated as daily average values (or hourly average values) by removing zero values from the data obtained from preterm infant vital signs after birth every minute. It is. Onset is the finding of retinopathy of prematurity (whether or not retinopathy of prematurity has developed) by a doctor, which is performed a predetermined number of times after birth. A count is obtained by converting the number of days elapsed from the date of birth into a unit time. The difference is a difference feature quantity obtained by calculating the difference for each measurement of each parameter. In addition, the weight, height, and vital signs, which are the input data in this embodiment, use daily average values, but may be 1-minute average values to 2-day average values, preferably 1-hour average values to 1 day. It is an average value, more preferably an hourly average value or a daily average value (hereinafter the same). If the average value for more than 2 days is used as input data, the prediction accuracy will be poor. .
　また、深層学習により学習済モデル１０を生成する場合、学習用特徴量演算部３３が学習済モデル１０に内蔵されている。具体的には、学習済モデル１０内部（畳み込み層）で時系列特徴情報を用いて重み付けを行い、学習済モデル１０内部（アテンション機構）で過去の予測特徴量を用いて重み付けする。ここで、時系列特徴情報とは学習用早産児情報３４ａを時系列に並べて特徴量としたものであり、過去の予測特徴量とは、学習済モデル１０自身が予測した結果として過去の時系列特徴情報の重み付けを現在の予測に用いるために取り出した重み付き特徴量のことである。 Also, when the trained model 10 is generated by deep learning, the learning feature quantity calculation unit 33 is built into the trained model 10 . Specifically, weighting is performed using the time-series feature information inside the trained model 10 (convolution layer), and weighting is performed using the past predicted feature amount inside the trained model 10 (attention mechanism). Here, the time-series feature information is a feature amount obtained by arranging the learning premature infant information 34a in time series, and the past prediction feature amount is the past time-series feature amount as a result predicted by the trained model 10 itself. It is a weighted feature extracted to use the weighting of feature information for the current prediction.
　このように生成された学習済モデル１０は、未熟児網膜症の治療要否のスコアを出力する。本実施形態における治療要否は、治療情報３４ｃと同様に、治療無し、治療有りType1 ROP、治療有りAPROPに区分され、治療要否のスコアは、治療無し、治療有りType1 ROP、治療有りAPROPにおける夫々の時系列のＡＵＣ (Area Under the Curve)で表される。この学習済モデル１０は、在胎週数が所定週（例えば２８週）未満である早産児の体重、身長及びバイタルサインに関する出生後の時系列データを含む早産児情報が入力されると、所定の生後日数（例えば２０日）以降で治療無し、治療有りType1 ROP、治療有りAPROP毎のスコアを出力することができる。この所定の生後日数は、１週間以上５週間以下であり、好ましくは、２週間以上４週間以下であり、更に好ましくは３週間近傍である（以下、同様）。スコアは、１日毎又は１時間毎の値（例えば時系列のＡＵＣ）として表される。本実施形態では、スコアを１日毎に算出しているが、スコアの算出間隔は、１分毎～２日毎、好ましくは１時間毎～１日毎、更に好ましくは１時間毎又は１日毎である（以下、同様）。スコアの算出間隔として、２日超である場合は予測精度が悪くなり、１分未満である場合はデータ量が多く、演算速度の低下やノイズの混入を招くおそれがある。このスコアは、治療無し、治療有りType1 ROP、治療有りAPROPの夫々に対して算出され、夫々のスコアのうち治療有りの値が最も高くなれば、近いうちに治療時間域に達する（治療適応である）。ここで、治療時間域とは、治療有りType1 ROPである場合、７２時間以内に治療を実施すること、又は、治療有りAPROPである場合、迅速に治療を実施することを言う。治療は、網膜光凝固、網膜冷凍凝固、抗VEGF薬の硝子体内投与、及び硝子体手術の何れかの方法から選択されるが、好ましくは、網膜光凝固又は抗VEGF薬の硝子体内投与である。硝子体手術は、網膜光凝固又は抗VEGF薬治療によっても効果が不十分で網膜剥離が発症したときに実施される。なお、学習済モデル１０は、将来的に治療を実施する（治療適応である）か否かをスコアとして表すだけでなく、治療実施時期又は治療時間域を出力するように構成しても良い。 The trained model 10 generated in this way outputs a score indicating the necessity of treatment for retinopathy of prematurity. Like the treatment information 34c, the need for treatment in this embodiment is classified into no treatment, Type 1 ROP with treatment, and APROP with treatment. Each time series is represented by AUC (Area Under the Curve). This trained model 10 is input with preterm infant information including postnatal chronological data on the weight, height and vital signs of a preterm infant whose gestational age is less than a predetermined week (for example, 28 weeks). It is possible to output scores for each of the postnatal days (for example, 20 days) without treatment, Type 1 ROP with treatment, and APROP with treatment. This predetermined number of days after birth is 1 week or more and 5 weeks or less, preferably 2 weeks or more and 4 weeks or less, and more preferably about 3 weeks (hereinafter the same). Scores are expressed as daily or hourly values (eg, AUC over time). In this embodiment, the score is calculated every day, but the score calculation interval is every minute to every two days, preferably every hour to every day, more preferably every hour or every day ( hereinafter the same). If the score calculation interval is more than two days, the prediction accuracy will be poor, and if it is less than one minute, the amount of data will be large, which may lead to a decrease in calculation speed and noise. This score is calculated for each of no treatment, Type 1 ROP with treatment, and APROP with treatment. If the score with treatment is the highest among each score, the treatment time range will be reached in the near future. be). Here, the treatment time window means that treatment should be performed within 72 hours in the case of Type 1 ROP with treatment, or that treatment should be promptly performed in the case of APROP with treatment. The treatment is selected from any of retinal photocoagulation, retinal cryocoagulation, intravitreal administration of an anti-VEGF drug, and vitrectomy, preferably retinal photocoagulation or intravitreal administration of an anti-VEGF drug. . Vitrectomy is performed when retinal detachment develops after retinal photocoagulation or anti-VEGF drug therapy is inadequate. Note that the learned model 10 may be configured to output a treatment implementation time or a treatment time range, in addition to expressing as a score whether or not treatment will be performed (treatment indicated) in the future.
　図２に示すように、スクリーニング装置４は、第二通信部４１、予測用特徴量演算部４２、リスク判定部４３、治療判定部４４、報知部４５及び第二記憶部４６を備えている。 As shown in FIG. 2, the screening device 4 includes a second communication unit 41, a prediction feature quantity calculation unit 42, a risk determination unit 43, a treatment determination unit 44, a notification unit 45, and a second storage unit 46.
　第二通信部４１は、インターネット回線２を介して、モニタリング機器１、学習済モデル生成装置３等との間でデータの送受信を行うインターフェースである。第二通信部４１は、モニタリング機器１から直接データを受信しても良いし、モニタリング機器１で取得されたデータをサーバ（不図示）に蓄積し、このサーバから蓄積されたデータを受信しても良い。 The second communication unit 41 is an interface that transmits and receives data to and from the monitoring device 1, the trained model generation device 3, etc. via the Internet line 2. The second communication unit 41 may receive data directly from the monitoring device 1, or accumulate data acquired by the monitoring device 1 in a server (not shown) and receive the accumulated data from this server. Also good.
　第二記憶部４６は、ＨＤＤやＳＳＤ等の一時的でない記憶媒体又はＲＡＭ等の一時的な記憶媒体で構成されており、プロセッサにより実行されるプログラムやアプリケーションを記憶している。第二記憶部４６は、第二通信部４１を介して取得したモニタリング機器１の予測用早産児情報４６ａ及び学習済モデル生成装置３が生成した学習済モデル１０を記憶している。予測用早産児情報４６ａは、在胎週数が所定週（例えば２８週）未満である早産児の体重、身長及びバイタルサインに関する出生後の時系列データを含んでいる。また、予測用早産児情報４６ａは、早産児の在胎日数及びアプガースコア等を含んでいる。この予測用早産児情報４６ａは、早産児の在胎週数及び治療時週数の合計が４０週以下の早産児から得た情報であることが好ましい。 The second storage unit 46 is composed of a non-temporary storage medium such as an HDD or SSD or a temporary storage medium such as a RAM, and stores programs and applications to be executed by the processor. The second storage unit 46 stores the predicted premature infant information 46 a of the monitoring device 1 acquired via the second communication unit 41 and the learned model 10 generated by the trained model generation device 3 . The predictive preterm infant information 46a includes postnatal chronological data on weight, height and vital signs of preterm infants whose gestational age is less than a predetermined week (eg, 28 weeks). The predictive premature infant information 46a also includes the gestational age and Apgar score of the premature infant. The predictive preterm infant information 46a is preferably information obtained from a preterm infant whose total gestational age and treatment weeks is 40 weeks or less.
　第二記憶部４６は、学習済モデル生成装置３で機械学習された学習済モデル１０に入力するために、予測用特徴量演算部４２により演算される予測用特徴量４６ｂを記憶している。 The second storage unit 46 stores the prediction feature quantity 46b calculated by the prediction feature quantity calculation unit 42 in order to input it to the learned model 10 machine-learned by the learned model generation device 3.
　第二記憶部４６は、学習済モデル１０から出力された判定結果４６ｃを記憶している。この判定結果４６ｃは、治療無し、治療有りType1 ROP、治療有りAPROPに区分された時系列データである。本実施形態における判定結果４６ｃは、治療無し、治療有りType1 ROP、治療有りAPROPにおけるＲＯＣ(Receiver Operating Characteristic) 曲線を含んでいる。また、判定結果４６ｃは、このＲＯＣ曲線から演算された時系列のＡＵＣ (Area Under the Curve)を含んでいる。 The second storage unit 46 stores the determination result 46c output from the trained model 10. This determination result 46c is time-series data divided into no treatment, Type 1 ROP with treatment, and APROP with treatment. The determination result 46c in this embodiment includes ROC (Receiver Operating Characteristic) curves for no treatment, Type 1 ROP with treatment, and APROP with treatment. The determination result 46c also includes a time-series AUC (Area Under the Curve) calculated from this ROC curve.
　予測用特徴量演算部４２は、プロセッサを備えており、予測用早産児情報４６ａから複数の予測用特徴量４６ｂを演算する。この予測用特徴量４６ｂは、発症を除いて学習用特徴量３４ｂと同様のものが用いられ、影響度の高い予測用特徴量４６ｂの順に、体重、身長、身長_ＳＤ、在胎日数、生後５分後のアプガースコア（Apgar Score; 5min）、体重_ＳＤ、生後１分後のアプガースコア（Apgar Score; 1min）、動脈血酸素飽和度（SpO2 %）、心拍数（HR bpm）、性別（M or F）、出生形態（単胎、双胎、品胎）、呼吸数（RESP/min）、呼吸数の差分（RESP/min.delta）、心拍数の差分（HR bpm.delta）、体重_ＳＤの差分、体重の差分、動脈血酸素飽和度の差分（SpO2 %.delta）、身長_ＳＤの差分、身長の差分となっている。 The prediction feature quantity calculation unit 42 includes a processor, and calculates a plurality of prediction feature quantities 46b from the prediction premature infant information 46a. This prediction feature quantity 46b is the same as the learning feature quantity 34b except for onset. Apgar score 5 minutes after birth (Apgar Score; 5 min), body weight SD, Apgar score 1 minute after birth (Apgar Score; 1 min), arterial blood oxygen saturation (SpO2 %), heart rate (HR bpm), gender (M or F ), birth type (single pregnancy, twin pregnancy, conception), respiratory rate (RESP/min), respiratory rate difference (RESP/min.delta), heart rate difference (HR bpm.delta), weight_SD difference, weight difference, arterial blood oxygen saturation difference (SpO2%.delta), height_SD difference, and height difference.
　リスク判定部４３は、プロセッサを備えており、所定の生後日数（例えば生後２０日）時点において、予測用早産児情報４６ａが入力された学習済モデル１０が未熟児網膜症の進行リスクを出力する。進行リスクは、１日毎又は１時間毎のスコアとして表される。このスコアは、治療無し、治療有りType1 ROP、治療有りAPROPの夫々に対して算出され、所定の生後日数時点において、夫々のスコアのうち治療有りの値が所定値よりも高ければ、進行リスク有りと判定する。一例として、リスク判定部４３は、治療無し、治療有りType1 ROP、治療有りAPROPで構成される複数の判定指標における時系列のＡＵＣから、所定の生後日数時点で最も高いＡＵＣを有する判定指標を抽出し、治療有りType1 ROP又は治療有りAPROPのＡＵＣが所定値（例えば０．３）よりも高ければ、進行リスク有りと判定する。この所定値は、０．１～０．８、好ましくは０．２～０．６、更に好ましくは０．３～０．５の間から設定される。 The risk determination unit 43 includes a processor, and the trained model 10 to which the predictive premature infant information 46a is input outputs the progression risk of retinopathy of prematurity at a predetermined number of days after birth (for example, 20 days after birth). . Progression risk is expressed as a daily or hourly score. This score is calculated for each of no treatment, Type 1 ROP with treatment, and APROP with treatment. If the value of each score with treatment is higher than a predetermined value at a predetermined number of days after birth, there is a risk of progression. I judge. As an example, the risk determination unit 43 extracts the determination index having the highest AUC at a predetermined number of days after birth from the time-series AUC of multiple determination indices composed of no treatment, type 1 ROP with treatment, and APROP with treatment. If the AUC of Type 1 ROP with treatment or APROP with treatment is higher than a predetermined value (for example, 0.3), it is determined that there is a risk of progression. This predetermined value is set between 0.1 and 0.8, preferably between 0.2 and 0.6, and more preferably between 0.3 and 0.5.
　治療判定部４４は、プロセッサを備えており、在胎週数が所定週（例えば２８週）未満である早産児の体重、身長及びバイタルサインに関する出生後の時系列データを含む予測用早産児情報４６ａが学習済モデル１０に入力されると、学習済モデル１０の出力値に基づいて所定の生後日数（例えば２０日）以降で未熟児網膜症の治療適応であるか否かを判定する。治療判定部４４は、リスク判定部４３で進行リスクが有ると判定された早産児のみ治療適応となるか否かを判定することが好ましい。未熟児網膜症の治療適応であるか否かは、１日毎又は１時間毎のスコアとして表される。このスコアは、治療無し、治療有りType1 ROP、治療有りAPROPの夫々に対して算出され、夫々のスコアのうち治療有りの値が最も高くなれば、治療判定部４４が近いうちに網膜光凝固等を用いた治療時間域に達する（治療適応となる）と判定する。一例として、治療判定部４４は、治療無し、治療有りType1 ROP、治療有りAPROPで構成される複数の判定指標における時系列のＡＵＣから、最も高いＡＵＣを有する判定指標を抽出し、治療有りType1 ROP又は治療有りAPROPのＡＵＣが治療無しのＡＵＣを上回れば、治療適応であると判定する。他の一例として、治療判定部４４は、治療無し、治療有りType1 ROP、治療有りAPROPで構成される複数の判定指標における時系列のＡＵＣから、治療有りType1 ROP又は治療有りAPROPのＡＵＣが治療閾値（例えば０．８）を上回れば、治療適応であると判定する。この治療閾値は、０．５～０．９、好ましくは０．６～０．９、更に好ましくは０．７～０．８の間から設定される。 The treatment determination unit 44 includes a processor, and predictive preterm infant information 46a including postnatal chronological data on weight, height, and vital signs of preterm infants whose gestational age is less than a predetermined number of weeks (for example, 28 weeks). is input to the trained model 10, it is determined based on the output value of the trained model 10 whether treatment for retinopathy of prematurity is indicated after a predetermined number of days after birth (for example, 20 days). It is preferable that the treatment determination unit 44 determines whether or not only premature infants determined by the risk determination unit 43 to have a risk of progression are indicated for treatment. Whether or not treatment for retinopathy of prematurity is indicated is expressed as a daily or hourly score. This score is calculated for each of no treatment, Type 1 ROP with treatment, and APROP with treatment. It is determined that the treatment time range using is reached (treatment is indicated). As an example, the treatment determination unit 44 extracts the determination index having the highest AUC from time-series AUCs in a plurality of determination indices including no treatment, Type 1 ROP with treatment, and APROP with treatment, and extracts the Type 1 ROP with treatment. Alternatively, if the AUC of APROP with treatment exceeds the AUC without treatment, it is determined that the patient is suitable for treatment. As another example, the treatment determination unit 44 determines the AUC of the Type 1 ROP with treatment or the APROP with treatment from the time-series AUC in a plurality of judgment indicators composed of no treatment, Type 1 ROP with treatment, and APROP with treatment. If it exceeds (for example, 0.8), it is determined that the treatment is indicated. This therapeutic threshold is set between 0.5 and 0.9, preferably between 0.6 and 0.9, more preferably between 0.7 and 0.8.
　報知部４５は、治療判定部４４が治療適応であると判定したとき、警告信号を出力する。報知部４５は、新生児集中治療室において早産児のバイタルサインを経時的にモニタリングする機器に搭載された警告ランプや警告音等で構成されても良いし、ナースステーションに備え付けられた所定の報知装置で構成されても良い。 The notification unit 45 outputs a warning signal when the treatment determination unit 44 determines that the treatment is indicated. The notification unit 45 may be composed of a warning lamp, a warning sound, or the like mounted on a device for chronologically monitoring the vital signs of premature infants in a neonatal intensive care unit, or a predetermined notification device provided at a nurse station. It may consist of
　続いて、図３から図１０を用いて、本実施形態に係る学習済モデル１０を用いて未熟児網膜症の進行予測を行うためにコンピュータが実行する未熟児網膜症スクリーニング方法（プログラム）の一例を説明する。 Next, with reference to FIGS. 3 to 10, an example of a screening method (program) for retinopathy of prematurity that is executed by a computer to predict the progression of retinopathy of prematurity using the trained model 10 according to the present embodiment. explain.
　学習済モデル生成装置３は、インターネット回線２を介して、夫々のモニタリング機器１から所定の期間に亘る学習用早産児情報３４ａ及び治療情報３４ｃを取得する（図３の＃３１）。図５に示すように、Ａ病院の早産児７１９人のデータに依ると、在胎週数が２７週になれば未熟児網膜症の発症確率が１０％程度に低下する。また、在胎週数が２８週未満の早産児のうち、出生体重が１０００ｇ未満の早産児はおよそ４０％が治療適応となり、出生体重が１０００ｇ未満の早産児が急速に悪化する未熟児網膜症（APROP）を発症するリスクが高い。このため、本実施形態では、在胎週数が２８週未満の早産児をスクリーニング対象とする。そこで、学習済モデル生成装置３は、学習用早産児情報３４ａ及び治療情報３４ｃとして、在胎週数が２８週未満の早産児に関するデータを抽出する（図３の＃３２、フィルタリング）。 The trained model generation device 3 acquires the learning preterm infant information 34a and the treatment information 34c over a predetermined period from each monitoring device 1 via the Internet line 2 (#31 in FIG. 3). As shown in FIG. 5, according to the data of 719 preterm infants in hospital A, the probability of developing retinopathy of prematurity decreases to about 10% when the gestational age reaches 27 weeks. In addition, among preterm infants with a gestational age of less than 28 weeks, approximately 40% of premature infants with a birth weight of less than 1000g are indicated for treatment, and premature infants with a birth weight of less than 1000g deteriorate rapidly (Retinopathy of prematurity). high risk of developing APROP). Therefore, in the present embodiment, premature infants whose gestational age is less than 28 weeks are targeted for screening. Therefore, the trained model generation device 3 extracts data on preterm infants with a gestational age of less than 28 weeks as the learning preterm infant information 34a and treatment information 34c (#32 in FIG. 3, filtering).
　図６の左図に示すように、Ａ病院にて治療を実施した早産児２０６人のデータに依ると、未熟児網膜症の治療時週数が生後６週目から１６週目が網膜光凝固等を用いた治療を実施する時期となる。また、図６の右図に示すように、Ａ病院にて治療を実施した早産児２０６人のデータに依ると、在胎週数及び治療時週数の合計が３０週目から３９週目が網膜光凝固等を用いた治療を実施する時期となる。このため、学習済モデル生成装置３は、学習用早産児情報３４ａ及び治療情報３４ｃとして、在胎週数及び治療時週数の合計が４０週以下の早産児に関するデータを抽出する（図３の＃３２、フィルタリング）。なお、学習済モデル生成装置３は、学習用早産児情報３４ａ及び治療情報３４ｃとして、在胎週数及び治療時週数の合計が２９週以上４０週以下の早産児に関するデータを抽出しても良い。また、学習済モデル生成装置３は、医師判断により早期治療した特異例となる、学習用早産児情報３４ａ及び治療情報３４ｃを除外する（図３の＃３２、フィルタリング）。その結果、本実施例における学習用早産児情報３４ａは、Ａ病院にて治療を実施した早産児２０６人における在胎日数、体重，身長，呼吸数，心拍数及び動脈血酸素飽和度の時系列データ、生後５分後のアプガースコア、生後１分後のアプガースコア、性別、出生形態、カウント、発症となっている（用語の定義は上述）。 As shown in the left figure of Fig. 6, according to the data of 206 premature infants who underwent treatment at hospital A, retinal photocoagulation was observed between the 6th and 16th weeks after birth for retinopathy of prematurity. It is time to implement treatment using In addition, as shown in the right figure of Fig. 6, according to the data of 206 preterm infants who underwent treatment at A hospital, the total number of gestational weeks and treatment weeks was 30 to 39 weeks. It is time to implement treatment using photocoagulation. For this reason, the trained model generation device 3 extracts, as the learning preterm infant information 34a and treatment information 34c, data relating to preterm infants whose total gestational age and treatment weeks are 40 weeks or less (# in FIG. 3). 32, filtering). Note that the trained model generation device 3 may extract, as the learning premature infant information 34a and the treatment information 34c, data on premature infants whose total gestational age and number of weeks at the time of treatment is 29 weeks or more and 40 weeks or less. . In addition, the trained model generating device 3 excludes the premature infant information for learning 34a and the treatment information 34c, which are peculiar cases of early treatment determined by the doctor (#32 in FIG. 3, filtering). As a result, the learning premature infant information 34a in this embodiment is time-series data of gestational age, weight, height, respiratory rate, heart rate, and arterial blood oxygen saturation for 206 preterm infants who underwent treatment at A hospital. , Apgar score at 5 minutes after birth, Apgar score at 1 minute after birth, sex, birth pattern, count, and onset (the terms are defined above).
　次いで、学習済モデル生成装置３のモデル生成部３２は、入力データが学習用早産児情報３４ａ（早産児の体重、身長及びバイタルサインに関する出生後の時系列データ等）で、教師データが治療情報３４ｃ（治療無し、治療有りType1 ROP、治療有りAPROP）として強化学習させ、学習済モデル１０を生成する（図３の＃３３～＃３６）。モデル生成部３２は、機械学習を行う場合（図３の＃３３Ｙｅｓ）、学習用特徴量演算部３３が学習用早産児情報３４ａから複数の学習用特徴量３４ｂを演算する（図３の＃３４、特徴量演算工程）。この学習用特徴量３４ｂは、在胎日数、体重の１日平均値、体重の差分、体重_ＳＤ、体重_ＳＤの差分、身長の１日平均値、身長の差分、身長_ＳＤ、身長_ＳＤの差分、呼吸数の１日平均値、呼吸数の差分、心拍数の１日平均値、心拍数の差分、動脈血酸素飽和度の１日平均値、動脈血酸素飽和度の差分、生後５分後のアプガースコア、生後１分後のアプガースコア、性別、出生形態、カウント、発症である（用語の定義は上述）。図８に示すように、Ａ病院にて治療を実施した早産児２０６人の学習用早産児情報３４ａを用いて演算された複数の学習用特徴量３４ｂを学習済モデル１０に入力して解析すると、各特徴量の治療情報３４ｃへの影響度が求められる。そして、モデル生成部３２は、入力データが学習用特徴量３４ｂで、教師データが治療情報３４ｃ（治療無し、治療有りType1 ROP、治療有りAPROP）として強化学習させ、学習済モデル１０を生成する（図３の＃３６）。 Next, the model generation unit 32 of the trained model generation device 3 receives the premature infant information for learning 34a as the input data (such as time-series data after birth regarding the weight, height and vital signs of the premature infant), and the teacher data as treatment information. 34c (without treatment, Type 1 ROP with treatment, APROP with treatment) and perform reinforcement learning to generate a learned model 10 (#33 to #36 in FIG. 3). When the model generation unit 32 performs machine learning (#33 Yes in FIG. 3), the learning feature amount calculation unit 33 calculates a plurality of learning feature amounts 34b from the learning premature infant information 34a (#34 in FIG. 3). , feature quantity calculation step). This learning feature value 34b includes the number of days of gestation, daily average weight, weight difference, weight_SD, weight_SD difference, daily average height, height difference, height_SD, height_ SD difference, daily mean respiratory rate, respiratory rate difference, daily mean heart rate, heart rate difference, daily mean arterial oxygen saturation, arterial blood oxygen saturation difference, 5 minutes after birth postnatal Apgar score, 1 minute postnatal Apgar score, sex, birth morphology, count, onset (defined above). As shown in FIG. 8, when a plurality of learning feature values 34b calculated using the learning premature infant information 34a of 206 premature infants treated at A hospital are input to the learned model 10 and analyzed, , the degree of influence of each feature amount on the treatment information 34c is obtained. Then, the model generation unit 32 performs reinforcement learning using the input data as the learning feature amount 34b and the teacher data as the treatment information 34c (no treatment, type 1 ROP with treatment, APROP with treatment), and generates the learned model 10 ( #36 in FIG. 3).
　一方、モデル生成部３２は、機械学習を行わない場合（図３の＃３３Ｎｏ）、畳み込みニューラルネットワークを含む深層学習を行う（図３の＃３５）。この深層学習において、モデル生成部３２は、入力データが学習用早産児情報３４ａ（早産児の体重、身長及びバイタルサインに関する出生後の時系列データ等）で、教師データが治療情報３４ｃ（治療無し、治療有りType1 ROP、治療有りAPROP）として強化学習させ、学習済モデル１０を生成する（図３の＃３６）。 On the other hand, when the model generation unit 32 does not perform machine learning (#33 No in FIG. 3), it performs deep learning including a convolutional neural network (#35 in FIG. 3). In this deep learning, the model generating unit 32 uses premature infant information 34a for learning as input data (postnatal time-series data on the weight, height and vital signs of premature infants, etc.) and teacher data as treatment information 34c (without treatment). , Type 1 ROP with treatment, APROP with treatment) to generate a learned model 10 (#36 in FIG. 3).
　次いで、スクリーニング装置４は、生後２０日時点において、予測用早産児情報４６ａが入力された学習済モデル１０が１日毎のスコア（ＲＯＣ曲線から演算された時系列のＡＵＣ）を出力し（図７参照）、リスク判定部４３がType1 ROP又はAPROPのスコアに基づいて未熟児網膜症の進行リスクを判定する（図３の＃３７、リスク判定工程）。この予測用早産児情報４６ａは、在胎日数、体重，身長，呼吸数，心拍数及び動脈血酸素飽和度の時系列データ、生後５分後のアプガースコア、生後１分後のアプガースコア、性別、出生形態、カウントとなっている（用語の定義は上述）。このスコアは、治療無し、治療有りType1 ROP、治療有りAPROPの夫々に対して算出され、生後２０日時点において、夫々のスコアのうち治療有りの値が所定値（例えば０．３）よりも高ければ、進行リスク有りと判定する（図３の＃３８Ｙｅｓ、図７の左図）。一方、夫々のスコアのうち治療有りの値が所定値以下であれば、進行リスク無しと判定する（図３の＃３８Ｎｏ、図７の右図）。 Next, the screening device 4 outputs the daily score (time-series AUC calculated from the ROC curve) of the trained model 10 to which the predictive premature infant information 46a is input at 20 days after birth (Fig. 7 See), the risk determination unit 43 determines the risk of progression of retinopathy of prematurity based on the score of Type 1 ROP or APROP (#37 in FIG. 3, risk determination step). The predictive premature infant information 46a includes time-series data of gestational age, weight, height, respiratory rate, heart rate, and arterial blood oxygen saturation, Apgar score 5 minutes after birth, Apgar score 1 minute after birth, sex, and birth pattern. , counts (the terms are defined above). This score is calculated for each of no treatment, Type 1 ROP with treatment, and APROP with treatment. If so, it is determined that there is a risk of progression (#38 Yes in FIG. 3, left diagram in FIG. 7). On the other hand, if the value of the presence of treatment among the respective scores is equal to or less than a predetermined value, it is determined that there is no progression risk (#38 No in FIG. 3, right diagram in FIG. 7).
　次いで、図４に示すように、進行リスク有りと判定された予測用早産児情報４６ａをオンセットデータ（Onset）とし、治療判定部４４は、オンセットデータである予測用早産児情報４６ａを学習済モデル１０に入力し、学習済モデル１０の出力値に基づいて生後２０日以降で未熟児網膜症の治療適応となるか否かを判定する（図３の＃３９、治療判定工程）。本実施形態における学習済モデル１０は、進行リスク有りと判定された予測用早産児情報４６ａの中でも、自然に治癒するケース（Spontaneous regression）と治療適応のケース（Disease progression）とを判別することができる。なお、機械学習した学習済モデル１０の場合、治療判定部４４は、在胎日数、体重の１日平均値、体重の差分、体重_ＳＤ、体重_ＳＤの差分、身長の１日平均値、身長の差分、身長_ＳＤ、身長_ＳＤの差分、呼吸数の１日平均値、呼吸数の差分、心拍数の１日平均値、心拍数の差分、動脈血酸素飽和度の１日平均値、動脈血酸素飽和度の差分、生後５分後のアプガースコア、生後１分後のアプガースコア、性別、出生形態、カウントで構成される予測用特徴量４６ｂを学習済モデル１０に入力する。 Next, as shown in FIG. 4, the predictive preterm infant information 46a determined to have progression risk is used as onset data (Onset), and the treatment determination unit 44 learns the predictive preterm infant information 46a, which is the onset data. Then, based on the output values of the trained model 10, it is determined whether or not treatment for retinopathy of prematurity is indicated after 20 days of birth (#39 in FIG. 3, treatment determination step). The trained model 10 in the present embodiment can distinguish between a spontaneously cured case (spontaneous regression) and a treatment adaptation case (disease progression) among the predictive premature infant information 46a determined to have a progression risk. can. In the case of the machine-learned model 10, the treatment determination unit 44 includes the number of gestational days, daily average weight, weight difference, weight_SD, weight_SD difference, daily average height, height difference, height_SD, height_SD difference, daily average respiratory rate, respiratory rate difference, daily average heart rate, heart rate difference, daily average arterial blood oxygen saturation, Prediction feature values 46b composed of difference in arterial blood oxygen saturation, Apgar score 5 minutes after birth, Apgar score 1 minute after birth, sex, birth pattern, and count are input to the trained model 10 .
　治療判定部４４は、リスク判定部４３で進行リスクが有ると判定された早産児に関する予測用早産児情報４６ａが学習済モデル１０に入力されると、学習済モデル１０の出力値に基づいて生後２０日以降で未熟児網膜症の治療適応となる否かを判定する。詳細に述べると、治療判定部４４は、治療無し、治療有りType1 ROP、治療有りAPROPの夫々に対するスコアのうち治療有りの値が最も高くなれば、治療判定部４４が近いうちに網膜光凝固等を用いた治療時間域に達する（治療適応である）と判定し（図３の＃４０Ｙｅｓ）、報知部４５が所定の手段により報知する（図３の＃４１）。図７の左図の例では、生後３週間程度で治療有りAPROPのスコア（ＡＵＣ）が最も高くなったため、治療適応であると判定する。 When the predictive premature infant information 46a regarding the premature infant determined by the risk determination unit 43 to have a progression risk is input to the learned model 10, the treatment determination unit 44 performs postnatal treatment based on the output value of the learned model 10. After 20 days, it is determined whether or not the treatment for retinopathy of prematurity is indicated. More specifically, the treatment determination unit 44 determines that if the value of the treatment with treatment among the scores for no treatment, type 1 ROP with treatment, and APROP with treatment is the highest, the treatment determination unit 44 will predict that retinal photocoagulation will occur in the near future. (#40 Yes in FIG. 3), and the notification unit 45 notifies by a predetermined means (#41 in FIG. 3). In the example on the left side of FIG. 7, the score (AUC) of APROP with treatment became the highest at about 3 weeks after birth, so it is determined that the treatment is indicated.
　図９には、Ａ病院（早産児２０６人）において上述した学習用特徴量３４ｂを入力データとして機械学習を行った学習済モデル１０を用いた未熟児網膜症の進行予測性能が示されており、図１０には、Ａ病院（早産児２０６人）において上述した学習用早産児情報３４ａを入力データとして深層学習を行った学習済モデル１０を用いた未熟児網膜症の進行予測性能が示されている。図９に示す検証データは、上述した予測用特徴量４６ｂを学習済モデル１０に入力し、未熟児網膜症の進行予測性能を、治療無し、治療有りType1 ROP、治療有りAPROPの夫々に対するスコア（ＲＯＣ曲線から演算された時系列のＡＵＣ）として表した。図１０に示す検証データは、上述した予測用早産児情報４６ａを学習済モデル１０に入力し、未熟児網膜症の進行予測性能を治療無し、治療有りType1 ROP、治療有りAPROPの夫々に対するスコア（ＲＯＣ曲線から演算された時系列のＡＵＣ）として表した。
図９の上段は、Ａ病院（早産児２０６人）の学習用特徴量３４ｂで機械学習した学習済モデル１０に、Ａ病院（早産児２０６人）の予測用特徴量４６ｂを入力した生後２０日目の進行予測性能（ＲＯＣ曲線）を示し、図９の下段は、Ａ病院（早産児２０６人）の学習用特徴量３４ｂで機械学習した学習済モデル１０に、Ｂ病院（早産児５９人）の予測用特徴量４６ｂを入力した生後２０日目の進行予測性能（ＲＯＣ曲線）を示している。図９の上段に示すように、Ａ病院における治療無しのＲＯＣ曲線下面積（ＡＵＣ）は０．６９であり、治療有りAPROPのＡＵＣは０．８２であり、治療有りType1 ROPのＡＵＣは０．５８であった。これらの結果から、進行例の未熟児網膜症について精度良く進行予測できることが分かる。また、図９の下段に示すように、Ｂ病院における治療無しのＲＯＣ曲線下面積（ＡＵＣ）は０．６６であり、治療有りAPROPのＡＵＣは０．８３であり、治療有りType1 ROPのＡＵＣは０．５８であり、Ａ病院における進行予測性能とほぼ同じであった。これより、Ａ病院の学習用特徴量３４ｂで機械学習した学習済モデル１０は、Ｂ病院における未熟児網膜症の進行予測ができる汎用性の高いモデルとなっている。図１０は、Ａ病院（早産児２０６人）の学習用早産児情報３４ａで深層学習した学習済モデル１０に、Ｂ病院（早産児５９人）の予測用早産児情報４６ａを入力した進行予測性能（ＡＵＣの時系列データ）を示しており、治療又は退院となった日から逆算した時系列予測性能である。同図に示すように、少なくとも治療５０日以上前にはＡＵＣが０．８以上となることから、未熟児網膜症が治療適応に進行する可能性のある症例を、時期を逸することなく判定できていることが分かる。これより、Ａ病院の学習用早産児情報３４ａで深層学習した学習済モデル１０は、Ｂ病院における未熟児網膜症の進行予測ができる汎用性の高いモデルとなっている。このように、本実施形態は、既存のモデル（WINROPやCHOP-ROP model）に比べて高い精度で未熟児網膜症の進行を予測できる。既存のモデルを種々の国であてはめた報告があるが、国によって精度に著しいばらつきがあることが分かっている。これは、新生児管理の医療水準の差のよるばらつきであると推測されている。本実施形態の学習済モデル１０は、新生児管理の異なる施設でも対応することができる。その結果、高度に専門的な知識と経験を有する者でなくとも治療適応を判断し、適切な時期に治療を開始することを可能にする。 FIG. 9 shows the progress prediction performance of retinopathy of prematurity using the trained model 10 that has undergone machine learning using the above-described learning feature value 34b as input data at Hospital A (206 premature infants). FIG. 10 shows the progress prediction performance of retinopathy of prematurity using a trained model 10 that has undergone deep learning using the above-described learning premature infant information 34a as input data at Hospital A (206 premature infants). ing. The verification data shown in FIG. 9 is obtained by inputting the above-described prediction feature value 46b into the trained model 10, and evaluating the progress prediction performance of retinopathy of prematurity as a score ( Time-series AUC calculated from the ROC curve). The verification data shown in FIG. 10 is obtained by inputting the above-described predictive premature infant information 46a into the learned model 10, and evaluating the performance of predicting the progression of retinopathy of prematurity with scores ( Time-series AUC calculated from the ROC curve).
The upper part of FIG. 9 is 20 days after birth when the prediction feature value 46b of hospital A (206 premature babies) is input to the trained model 10 machine-learned with the learning feature value 34b of hospital A (206 premature babies). The progression prediction performance (ROC curve) of the eye is shown, and the lower part of FIG. 20 shows the progress prediction performance (ROC curve) on the 20th day after birth when the prediction feature value 46b is input. As shown in the upper part of FIG. 9, the area under the ROC curve (AUC) without treatment in A hospital is 0.69, the AUC of APROP with treatment is 0.82, and the AUC of Type 1 ROP with treatment is 0.69. was 58. These results show that the progression of retinopathy of prematurity can be predicted with high accuracy. In addition, as shown in the lower part of FIG. 9, the area under the ROC curve (AUC) without treatment at B hospital is 0.66, the AUC of APROP with treatment is 0.83, and the AUC of Type 1 ROP with treatment is It was 0.58, which was almost the same as the progression prediction performance in A hospital. Thus, the trained model 10 machine-learned with the learning feature quantity 34b of the A hospital is a highly versatile model capable of predicting the progression of retinopathy of prematurity in the B hospital. FIG. 10 shows the progress prediction performance when the premature infant information 46a for prediction of hospital B (59 preterm infants) is input to the trained model 10 that has undergone deep learning with the premature infant information 34a for learning of hospital A (206 preterm infants). (time-series data of AUC), which is the time-series prediction performance calculated backward from the date of treatment or discharge. As shown in the figure, since the AUC is 0.8 or more at least 50 days before treatment, cases in which retinopathy of prematurity may progress to treatment indications can be determined in a timely manner. I know it's done. As a result, the trained model 10 that has undergone deep learning with the learning premature infant information 34a of A hospital is a highly versatile model capable of predicting the progression of retinopathy of prematurity at B hospital. Thus, the present embodiment can predict the progression of retinopathy of prematurity with higher accuracy than existing models (WINROP and CHOP-ROP models). There are reports of fitting existing models in various countries, but we know that the accuracy varies significantly from country to country. It is speculated that this is due to differences in the medical standards of neonatal care. The trained model 10 of this embodiment can also be used in facilities that manage different newborns. As a result, even those who do not have highly specialized knowledge and experience can judge treatment indications and start treatment at an appropriate time.
　本実施形態における学習済モデル１０は、適切なタイミングで未熟児網膜症の進行予測を精度良く行うことが可能な、汎用性の高いものである。また、在胎週数及び治療時週数の合計が４０週より大きければ、未熟児網膜症の発症リスクが極めて小さいため、在胎週数及び治療時週数の合計が４０週以下の早産児情報を用いて学習した、本実施形態における学習済モデル１０は、未熟児網膜症の進行予測を精度良く行うことができる。 The trained model 10 in the present embodiment is highly versatile, capable of accurately predicting the progression of retinopathy of prematurity at an appropriate timing. In addition, if the total number of weeks of gestation and the number of weeks at the time of treatment is greater than 40 weeks, the risk of developing retinopathy of prematurity is extremely low. The trained model 10 according to the present embodiment, which is learned using the method, can accurately predict the progression of retinopathy of prematurity.
　また、治療無し、治療有りType1 ROP、治療有りAPROP毎のスコアを出力する学習済モデル１０を用いて所定の生後日数以降で未熟児網膜症の治療適応となるか否かを判定することにより、未熟児網膜症の進行を予測することができる。未熟児網膜症の発症から治療適応となるまでの進行過程には多くの要素が複雑に関連し、その要素は経時的に変動するが、本実施形態では、複数の時系列データを用いて未熟児網膜症の進行を高い精度で予測する方法となっている。しかも、本実施形態では、学習済モデル１０が未熟児網膜症の進行リスクを出力するため、未熟児網膜症の潜在的な進行度合いを推定することができる。そして、進行リスクが有ると判定された早産児のみ治療適応となるか否かを識別するため、２段階の判定工程により精度の高い未熟児網膜症スクリーニング方法となっている。 In addition, by determining whether or not treatment for retinopathy of prematurity is indicated after a predetermined number of days after birth using a trained model 10 that outputs scores for each type 1 ROP with treatment, type 1 ROP with treatment, and APROP with treatment, It can predict the progression of retinopathy of prematurity. Many elements are intricately related to the progression process from the onset of retinopathy of prematurity to the time when treatment is indicated, and the elements fluctuate over time. It is a method for predicting the progression of infantile retinopathy with high accuracy. Moreover, in the present embodiment, since the learned model 10 outputs the risk of progression of retinopathy of prematurity, the latent degree of progression of retinopathy of prematurity can be estimated. In order to identify whether or not only premature infants determined to have a risk of progression are eligible for treatment, the two-step determination process provides a highly accurate screening method for retinopathy of prematurity.
［その他の実施形態］
（１）学習済モデル１０が未熟児網膜症の進行リスクを出力するリスク判定工程を省略しても良い。この場合でも、治療無し、治療有りType1 ROP、治療有りAPROP毎のスコアを出力する学習済モデル１０を用いて所定の生後日数以降で未熟児網膜症の治療適応となるか否かを判定する治療判定工程により、未熟児網膜症の進行予測を精度良く行うことができる。
（２）学習済モデル１０は、決定木以外で構成される機械学習や、畳み込みニューラルネットワーク以外の深層学習によって生成しても良い。例えば、サポートベクターマシンやロジスティック回帰等の公知の学習手法を用いることができる。
（３）早産児情報は、体重、身長及びバイタルサインに関する出生後の時系列データを含むものであれば、その他のパラメータを用いることができる。 [Other embodiments]
(1) The risk determination step in which the trained model 10 outputs the progression risk of retinopathy of prematurity may be omitted. Even in this case, treatment to determine whether or not retinopathy of prematurity is indicated for treatment after a predetermined number of days after birth using a trained model 10 that outputs scores for each of no treatment, Type 1 ROP with treatment, and APROP with treatment. The determination step can accurately predict the progression of retinopathy of prematurity.
(2) The trained model 10 may be generated by machine learning other than a decision tree, or by deep learning other than a convolutional neural network. For example, a known learning method such as support vector machine or logistic regression can be used.
(3) Premature infant information can include other parameters as long as they include postnatal chronological data on weight, height and vital signs.
　本開示は、未熟児網膜症の進行予測を行う未熟児網膜症スクリーニング方法、スクリーニング装置及び学習済モデルに利用可能である。 The present disclosure can be used for a retinopathy of prematurity screening method, screening device, and trained model for predicting the progression of retinopathy of prematurity.
４　　　：スクリーニング装置
１０　　：学習済モデル
３４ａ　：学習用早産児情報（早産児情報）
３４ｂ　：学習用特徴量（特徴量）
４６ａ　：予測用早産児情報（早産児情報）
  4: Screening device 10: Trained model 34a: Premature infant information for learning (premature infant information)
34b: feature quantity for learning (feature quantity)
46a: Preterm infant information for prediction (preterm infant information)

Claims (12)

  1. 　未熟児網膜症の進行予測を行う未熟児網膜症スクリーニング方法であって、
    　在胎週数が所定週未満である早産児の体重、身長及びバイタルサインに関する出生後の時系列データを含む早産児情報に基づいて、所定の生後日数以降で未熟児網膜症の治療適応となるか否かを判定する治療判定工程を含む未熟児網膜症スクリーニング方法。     A screening method for retinopathy of prematurity for predicting the progression of retinopathy of prematurity, comprising:
    Is treatment indicated for retinopathy of prematurity after a given number of days after birth based on preterm infant information, including postnatal chronological data on weight, height, and vital signs for preterm infants with a gestational age of less than a given number of weeks? A screening method for retinopathy of prematurity, comprising a treatment determination step of determining whether or not
  2. 　前記所定の生後日数時点において算出されたType1 ROP又はAPROPのスコアに基づいて未熟児網膜症の進行リスクを判定するリスク判定工程を更に含み、
    　前記リスク判定工程で前記進行リスクが有ると判定された前記早産児のみ前記治療判定工程を実行する請求項１に記載の未熟児網膜症スクリーニング方法。     Further comprising a risk determination step of determining the risk of progression of retinopathy of prematurity based on the Type 1 ROP or APROP score calculated at the predetermined number of days after birth;
    2. The method of screening for retinopathy of prematurity according to claim 1, wherein the treatment determination step is performed only for the premature infant determined to have the progression risk in the risk determination step.
  3. 　前記バイタルサインは、前記早産児の心拍数、呼吸数及び動脈血酸素飽和度の少なくとも１つである請求項１又は２に記載の未熟児網膜症スクリーニング方法。 The method for screening retinopathy of prematurity according to claim 1 or 2, wherein the vital sign is at least one of heart rate, respiratory rate and arterial blood oxygen saturation of the premature infant.
  4. 　前記早産児情報は、前記早産児の在胎日数及びアプガースコアの少なくとも１つを含んでいる請求項１から３の何れか一項に記載の未熟児網膜症スクリーニング方法。 The method of screening for retinopathy of prematurity according to any one of claims 1 to 3, wherein the preterm infant information includes at least one of the gestational age and Apgar score of the preterm infant.
  5. 　未熟児網膜症の進行予測を行うスクリーニング装置であって、
    　在胎週数が所定週未満である早産児の体重、身長及びバイタルサインに関する出生後の時系列データを含む早産児情報に基づいて、所定の生後日数以降で未熟児網膜症の治療適応となるか否かを判定する治療判定部を備えたスクリーニング装置。     A screening device for predicting the progression of retinopathy of prematurity,
    Is treatment indicated for retinopathy of prematurity after a given number of days after birth based on preterm infant information, including postnatal chronological data on weight, height, and vital signs for preterm infants with a gestational age of less than a given number of weeks? A screening device comprising a treatment determination unit that determines whether or not.
  6. 　前記所定の生後日数時点において算出されたType1 ROP又はAPROPのスコアに基づいて未熟児網膜症の進行リスクを判定するリスク判定部を更に備え、
    　前記治療判定部は、前記リスク判定部で前記進行リスクが有ると判定された前記早産児のみを対象として前記治療適応となるか否かを判定する請求項５に記載のスクリーニング装置。     Further comprising a risk determination unit that determines the risk of progression of retinopathy of prematurity based on the Type 1 ROP or APROP score calculated at the predetermined number of days after birth,
    The screening device according to claim 5, wherein the treatment determination unit determines whether or not the treatment is indicated for only the premature infant determined by the risk determination unit to have the risk of progression.
  7. 　コンピュータにより機能する学習済モデルであって、
    　ツリー構造に並んだ複数の分岐点からなる決定木で構成され、
    　未熟児網膜症の治療を行った在胎週数が所定週未満である早産児の体重、身長及びバイタルサインに関する出生後の時系列データを含む早産児情報に基づいて演算した特徴量が入力され、夫々の前記分岐点での評価値を合算することにより、未熟児網膜症の治療要否のスコアを出力する学習済モデル。     A trained model functioning with a computer, comprising:
    It consists of a decision tree consisting of multiple branch points arranged in a tree structure,
    A feature value calculated based on preterm infant information including postnatal chronological data on weight, height, and vital signs of a preterm infant whose gestational age is less than a predetermined week and who has been treated for retinopathy of prematurity is input, A trained model that outputs a score indicating whether or not treatment is necessary for retinopathy of prematurity by summing the evaluation values at each branch point.
  8. 　コンピュータにより機能する学習済モデルであって、
    　畳み込みニューラルネットワークを含む深層学習で生成され、
    　未熟児網膜症の治療を行った在胎週数が所定週未満である早産児の体重、身長及びバイタルサインに関する出生後の時系列データを含む早産児情報が入力され、未熟児網膜症の治療要否のスコアを出力する学習済モデル。     A trained model functioning with a computer, comprising:
    generated by deep learning, including convolutional neural networks,
    Premature infant information including chronological postnatal data on weight, height and vital signs of a premature infant with a gestational age of less than a predetermined number of weeks who received treatment for retinopathy of prematurity is entered, and a treatment required for retinopathy of prematurity is entered. A trained model that outputs a negative score.
  9. 　前記早産児情報は、前記早産児の在胎週数及び治療時週数の合計が４０週以下の前記早産児から得た情報である請求項７又は８に記載の学習済モデル。 The learned model according to claim 7 or 8, wherein the premature infant information is information obtained from the premature infant whose total gestational age and number of weeks at the time of treatment is 40 weeks or less.
  10. 　前記早産児情報は、医師判断により早期治療した前記早産児から得た情報を除外したものである請求項７から９の何れか一項に記載の学習済モデル。 The learned model according to any one of claims 7 to 9, wherein the premature infant information excludes information obtained from the premature infant who was treated at an early stage by doctor's judgment.
  11. 　前記バイタルサインは、前記早産児の心拍数、呼吸数及び動脈血酸素飽和度の少なくとも１つである請求項７から１０の何れか一項に記載の学習済モデル。 The learned model according to any one of claims 7 to 10, wherein the vital signs are at least one of heart rate, respiratory rate and arterial blood oxygen saturation of the premature infant.
  12. 　前記早産児情報は、前記早産児の在胎日数及びアプガースコアの少なくとも１つを含んでいる請求項７から１１の何れか一項に記載の学習済モデル。
          12. The trained model according to any one of claims 7 to 11, wherein the preterm infant information includes at least one of gestational age and Apgar score of the preterm infant.
PCT/JP2022/022042 2021-06-03 2022-05-31 Method for screening for retinopathy of prematurity, apparatus for screening for retinopathy of prematurity, and learned model WO2022255332A1 (en)

Priority Applications (2)

Application Number Priority Date Filing Date Title
US18/566,168 US20240274295A1 (en) 2021-06-03 2022-05-31 Method for screening for retinopathy of prematurity, apparatus for screening for retinopathy of prematurity, and learned model
JP2023525837A JPWO2022255332A1 (en) 2021-06-03 2022-05-31

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP2021-093588 2021-06-03
JP2021093588 2021-06-03

Publications (1)

Publication Number Publication Date
WO2022255332A1 true WO2022255332A1 (en) 2022-12-08

Family

ID=84323264

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2022/022042 WO2022255332A1 (en) 2021-06-03 2022-05-31 Method for screening for retinopathy of prematurity, apparatus for screening for retinopathy of prematurity, and learned model

Country Status (3)

Country Link
US (1) US20240274295A1 (en)
JP (1) JPWO2022255332A1 (en)
WO (1) WO2022255332A1 (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110295093A1 (en) * 2010-05-28 2011-12-01 Nellcor Puritan Bennett Llc Retinopathy Of Prematurity Determination And Alarm System
JP2013536971A (en) * 2010-09-07 2013-09-26 ザ ボード オブ トラスティーズ オブ ザ レランド スタンフォード ジュニア ユニバーシティー Medical scoring system and method
JP2015518491A (en) * 2012-05-04 2015-07-02 アクセラ インコーポレイテッド Methods for treating diabetic retinopathy and other eye diseases
US20180235467A1 (en) * 2015-08-20 2018-08-23 Ohio University Devices and Methods for Classifying Diabetic and Macular Degeneration

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110295093A1 (en) * 2010-05-28 2011-12-01 Nellcor Puritan Bennett Llc Retinopathy Of Prematurity Determination And Alarm System
JP2013536971A (en) * 2010-09-07 2013-09-26 ザ ボード オブ トラスティーズ オブ ザ レランド スタンフォード ジュニア ユニバーシティー Medical scoring system and method
JP2015518491A (en) * 2012-05-04 2015-07-02 アクセラ インコーポレイテッド Methods for treating diabetic retinopathy and other eye diseases
US20180235467A1 (en) * 2015-08-20 2018-08-23 Ohio University Devices and Methods for Classifying Diabetic and Macular Degeneration

Also Published As

Publication number Publication date
JPWO2022255332A1 (en) 2022-12-08
US20240274295A1 (en) 2024-08-15

Similar Documents

Publication Publication Date Title
Charlton et al. Wearable photoplethysmography for cardiovascular monitoring
Sloan et al. Vagally-mediated heart rate variability and indices of well-being: Results of a nationally representative study.
Flynn et al. Retinopathy of prematurity: diagnosis, severity, and natural history
Le Guen et al. The Analgesia Nociception Index: a pilot study to evaluation of a new pain parameter during labor
Sheinkopf et al. Developmental trajectories of autonomic functioning in autism from birth to early childhood
JPH10506034A (en) Fetal data processing system and method
Leganes-Fonteneau et al. Cardiovascular mechanisms of interoceptive awareness: Effects of resonance breathing
US20220175324A1 (en) Computer-based prediction of fetal and maternal outcomes
KR20210114580A (en) Wearable sleep monitoring system and method
Dell'Acqua et al. Depressed mood, brooding rumination and affective interference: The moderating role of heart rate variability
Wühl Ambulatory blood pressure monitoring methodology and norms in children
Cabrera et al. Combining ROPtool measurements of vascular tortuosity and width to quantify plus disease in retinopathy of prematurity
WO2022255332A1 (en) Method for screening for retinopathy of prematurity, apparatus for screening for retinopathy of prematurity, and learned model
JP6810135B2 (en) A system for detecting and early predicting the impending symptom deterioration of patients with chronic obstructive broncontomy
US20230200746A1 (en) System for acquisition and analysis of maternal and/or fetal physiological signals
CN113261924A (en) Intelligent stroke early warning system and method
JP7467033B2 (en) Medical information processing device
Zhivolupova Remote monitoring system for preeclampsia detection and control
AU2010239138A1 (en) A method and system for determining a variation in a metabolic function and managing the variation accordingly
Avrunin et al. Possibilities of apnea diagnostics by fuzzy logic methods
Berger et al. Neonatal rhinorrhea, heart rate variability, and childhood exercise-induced wheeze
Lee Cardiac vagal tone as an index of autonomic nervous function in healthy newborn and premature infants
Higgins Heart Rate Variability as an Assessment of Fall Risk in Older Adults
Gierałtowski et al. Formation of functional associations across time scales in the fetal autonomic control system—A multifractal analysis
Virkki et al. Overnight features of transcutaneous carbon dioxide measurement as predictors of metabolic status

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22816079

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 2023525837

Country of ref document: JP

WWE Wipo information: entry into national phase

Ref document number: 18566168

Country of ref document: US

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 22816079

Country of ref document: EP

Kind code of ref document: A1